# BaseModel.ai — Full content > BaseModel.ai is an enterprise-grade behavioral foundation model built by Synerise. It replaces the traditional ML stack (manual feature engineering plus a separate model per use case) with a single self-supervised model trained directly on raw event data from your data warehouse. From one foundation, lightweight "scenario heads" handle churn, LTV, recommendations, fraud detection, and more — going from raw events to production-ready predictions in hours or days instead of months. BaseModel.ai is deployed inside the customer's environment (Snowflake Native App, Databricks, or Docker on GPU) so behavioral data never leaves the customer's VPC. This file (llms-full.txt) contains the full readable body content of every page on basemodel.ai, concatenated as Markdown in reading order. Each page begins with a level-1 heading and its canonical source URL. For a shorter index with one-line descriptions, see https://basemodel.ai/llms.txt --- # Home — https://basemodel.ai/ ## Hero +138 % vs HSTU — New SOTA in sequential recommendation ### The behavioral foundation model. A single self-supervised model that ingests your entire data warehouse and turns any behavioral question about any individual into a production prediction — in hours, not months, without scaling your team. - Book a Demo - Explore Platform - **8 B+** — Events per training run - **18 M+** — Clients per training run - **<1 ms** — Classification latency - **9 ms** — 6 M-item recs per user (1× H100) Industries: Retail, Banking, Ecommerce, Telco, Travel, Gaming, Insurance, Automotive ## Why? ### Predicting behaviors is a key component of the future Humans have always built mental models of each other — predicting what someone might do next, what they need, what they fear. It's the foundation of civilization: trade, diplomacy, medicine, law. But these mental models don't scale. A doctor can know 2,000 patients. A sales rep can track 200 leads. A marketer can segment 20 audiences. The gap between what we *could* predict and what we *actually* predict is enormous. BaseModel.ai closes that gap. It ingests your entire data warehouse — years of clicks, purchases, sessions, and transactions across your whole population — and builds a living mathematical model of each individual. Not a summary. Not a segment. A continuously evolving behavioral fingerprint for every person. The result: predictions that no human team, and no language model, could ever produce. > "Every company sits on a goldmine of behavioral data, but they mine it one question at a time — one model for churn, another for conversion, another for fraud. We built BaseModel.ai so the number of behavioral predictions you can run is limited only by your imagination — not your headcount." Jack Dąbrowski — Chief AI Officer, Synerise / Founder, BaseModel.ai ## Common misconceptions ### Let's be clear. ### Not a language model BaseModel doesn't generate text. It generates predictions — scored, ranked, and ready for production serving at sub-millisecond latency. The output is a numeric score, not a paragraph. ### Not an AI coding assistant LLMs like ChatGPT and Claude can help a data scientist write an XGBoost pipeline. BaseModel replaces the need for that pipeline entirely. It has already learned your data. ### Not limited to a context window LLMs process thousands of tokens at a time. BaseModel ingests petabyte-scale data warehouses — years of behavioral history across millions of individuals — and retains all of it. ### Not a single-purpose model Traditional ML builds one model per question. BaseModel pre-trains once on all your behavioral data, then produces a fine-tuned predictor for any behavior you can define on an individual entity — whether that's purchase propensity, employee attrition, patient readmission, or account takeover. ## Understanding people at scale is hard Companies face fundamental challenges when trying to model human behavior with traditional approaches. ### Lack of time Data science teams spend 80% of their time on data preparation, not modeling. ### Evolving schemas Data schemas change constantly, breaking pipelines and invalidating features. ### Maintaining pipelines Every model needs its own ETL pipeline — a maintenance nightmare at scale. ### Handcrafting features Feature engineering requires deep domain expertise and months of iteration. ### High-dimensional spaces Human intuition fails in high-dimensional feature spaces — critical patterns go undetected. ### Different formats per model Each model expects different input formats, creating integration complexity. ### Guaranteeing correctness Validating feature correctness across hundreds of pipelines is nearly impossible. ## State of the Now ### The current workflow is broken Most organizations hit the same bottlenecks — and they compound. ### Manual labor & lack of time - Expert knowledge required for every model - Months of feature engineering per use case - Human limitations in understanding high-dimensional data ### Complicated data pipelines - Different ETL for every model - Multiple models with conflicting requirements - Spaghetti code that nobody dares to touch ### Organizational friction - Data science silos across departments - Data bias goes undetected - Complacent teams maintaining legacy systems #### The consequences - Hard cap on model complexity - Spaghetti code across teams - Data science silos - Missed revenue opportunities ## What BaseModel.ai does ### Reduce the modeling lifecycle from months to days ### Automatic feature creation from raw data No more manual feature engineering. BaseModel.ai learns directly from raw behavioral events — clicks, purchases, page views, app opens — and automatically creates the optimal feature representation. Eliminates - Manual feature engineering - Domain-expert bottleneck - Feature store maintenance ### Multi-source learning Combine data from any source — web, mobile, CRM, transactions, support tickets — into a single unified model. No data silos, no integration headaches. Eliminates - Data source silos - Format conversion pipelines - Missing cross-channel insights ### Fine-tuned models by defining an objective Define what you want to predict about any individual — whether it's purchase likelihood, attrition risk, or a question your competitors haven't even thought to ask — and the model fine-tunes automatically. No new pipeline. No new team. Eliminates - Model architecture decisions - Hyperparameter tuning - Training infrastructure setup ### Deepest explainability Every prediction comes with clear explanations. Understand which behaviors drove each decision — from individual predictions to aggregate insights. Eliminates - Black-box predictions - Compliance risk - Stakeholder trust issues ## Embedding Space Visualization BaseModel.ai projects every customer into a shared behavioral embedding space. Similar behaviors cluster together — enabling instant similarity search, segmentation, and anomaly detection. Clusters: High-Value Shoppers, At-Risk Churners, New Users, Power Users, Seasonal Buyers, Deal Seekers ### Behavioral Clustering Similar browsing, purchasing, and engagement patterns naturally cluster — no manual segmentation rules needed. ### Real-Time Drift Detection When a vector moves toward a different cluster, the system triggers proactive interventions automatically. ### Analogical Reasoning Vector arithmetic in embedding space reveals emerging high-value prospects and hidden segments. ### Cross-Entity Relationships Products, campaigns, and channels exist in the same space — enabling nearest-neighbor recommendations. ## BaseModel.ai vs. the Alternatives See how BaseModel.ai compares to traditional ML pipelines and the latest generation of LLM-powered AI tools. | Capability | Traditional ML | LLMs / AI Agents | BaseModel.ai | | --- | --- | --- | --- | | What it outputs | One model per business question | Text, code, or analysis | Production-ready predictive models | | Data scale | Curated datasets (GB) | Context window (128K–1M tokens) | Petabyte-scale data warehouses | | Feature engineering | Months of manual work per model | Can generate feature code — still single-purpose | Fully automated from raw events | | Time to first model | 3–6 months | Hours (but builds traditional pipelines) | 12 h foundation training on 1× A100 for ~8 B events; scenario fine-tune in hours | | Population modeling | Aggregate statistics and segments | One user at a time via prompts | Individual-level models for entire population | | Cross-domain transfer | Not possible — each model is siloed | Not applicable — no persistent learned state | Built-in — one model serves all domains | | Knowledge persistence | Retrain from scratch for each question | No memory between sessions | Foundation reused across every scenario | | Cold-start handling | Requires minimum data thresholds | Requires detailed prompt context | Inductive sketches from first interaction | | Team required | 5–15 ML engineers + data scientists | Data scientist + prompt engineer | Data engineer or ML engineer; a single ML engineer suffices for typical deployments | | Latency at scale | 50–500ms typical | 1–30 seconds per generation | Sub-ms classification; 9 ms/user to score 6 M items on 1× H100 | ## The Pre-Training & Fine-Tuning Pipeline From raw behavioral data to production-ready predictions — a four-stage pipeline that eliminates traditional ML complexity. ### PHASE 1 · Cleora-NX · Proprietary Hypergraph Embeddings Each behavioural event is a hyperedge linking everyone who took part — user, product, category, brand, timestamp — simultaneously. **Cleora-NX — Synerise's proprietary graph-embedding engine, building on the research published at ICONIP 2021. Production-only; with substantial performance and functionality extensions — not open-sourced.** Runs in time proportional to the number of hyperedges; converges in a handful of iterations on graphs with billions of interactions in minutes. - Cleora-NX — proprietary, never published; the open-source Cleora is its predecessor - Multi-modality support and temporal-interaction extensions on top of the public Cleora update - Deterministic — no training variance, no GPU required - Scales to graphs with millions of nodes and billions of interactions ### PHASE 2 · TREMDE · Proprietary Density Sketches Cleora-NX embeddings (plus any text, image, or tabular embeddings) are aggregated into fixed-size density sketches by **TREMDE — Synerise's proprietary, temporally-aware extension of the open-source EMDE algorithm (Synerise, ICONIP 2021)**. TREMDE produces **sparse codes that are composable under summation** — adding two entities' codes yields the code for their combined behaviour — and adds temporal-interaction modelling, modality-specific extensions, and other internal improvements **not described in the original EMDE paper**. Each modality is sketched independently and concatenated; the sketch shape itself is **auto-tuned by the pipeline from your data — no manual config, no hyperparameter sweep**. New items with content features get meaningful codes immediately — true cold-start with no retraining. - TREMDE — proprietary, never published; the open-source EMDE is its predecessor - Adds temporal-interaction modelling and per-modality extensions on top of public EMDE - Sketch shape auto-tuned from your data — zero manual config - Multimodal: graph, text, image, tabular features ### PHASE 3 · Foundation · Customized FFN Backbone Not a textbook MLP. A **customized feed-forward backbone** purpose-built for behavioural sequences, carrying proprietary inductive biases tuned to the structure of event data. **No attention, no recurrence** — yet it captures temporal structure, cross-feature interactions, and modality fusion that an off-the-shelf MLP cannot. The objective is distributional matching: each depth of the target sketch is normalized to sum to 1, and the loss is the cross-entropy between predicted and true depth distributions, averaged across depths and modalities. - Captures temporal structure without recurrence or attention - Models cross-feature interactions and modality fusion in one pass - Single foundation transfers across scenarios — no per-task retraining - Distributional cross-entropy across depths and modalities ### PHASE 4 · Serving · Proprietary Scoring · Ray Serve Candidate items are encoded into TREMDE sparse codes (composable under summation). Predictions are scored using a proprietary aggregation method optimized for behavioral sketches. Sub-millisecond classification latency; 9 ms per user to score the full 6 M-item rel-avito catalog on 1× H100. Served via Ray Serve. - Proprietary aggregation method optimized for behavioural sketches - Sub-millisecond latency for classification / regression - 9 ms per user to score 6 M-item catalog on 1× H100 - Ray Serve deployment for batch and online scoring ## Under a week: from data connection to production scoring A typical first deployment moves from data connection to production scoring on a representative dataset in under a week. Foundation training time scales with data volume — reference number: 12 hours on 1× NVIDIA A100 for ~8 B events / ~18 M clients. Each new scenario reuses the foundation via a quick fine-tune. Up to 3 days · Up to 2 days ### Deploy - Cloud or on-premise; Docker container - GPU: A100 recommended; multi-GPU A10s / L40s minimum (CUDA 12+) - 240 GB RAM, 32 CPU cores, 1 TB disk - Snowpark Container Services or your own Kubernetes ### Retrieve data Snowflake, BigQuery, Azure Synapse, Databricks, Hive, ClickHouse, Parquet Seven first-class connectors, one YAML data block. ### Foundation model training Self-supervised behavioral foundation model ### Fine tuning Fine-tuned supervised models for specific use-cases - Deploy + Retrieve + Train — Up to 3 days - Fine tuning — Up to 2 days ## Built different Six architectural principles that set BaseModel.ai apart from one-off ML pipelines and generic LLM tooling. ### Reusable Foundation The foundation model is trained once on your behavioural data. Every new scenario — churn, LTV, recommendations — reuses those embeddings via a quick fine-tune; no full retraining for each question. ### Multimodal Density Sketches Graph embeddings, text, images, and tabular features are sketched into a single fixed-size representation per entity — built on-the-fly with cost linear in interactions. ### Zero-Shot Predictions Answer behavioural questions you've never explicitly modelled — such as patient readmission risk, subscriber upgrade propensity, or employee flight risk — by defining a target function and reusing the foundation embeddings. ### Real-Time Inference Sub-millisecond classification and regression latency; recommendations score the full 6 M-item rel-avito catalog in 9 ms per user on 1× H100. Served via Ray Serve for batch and online workloads. ### Self-Hosted by Design Deploy inside Snowflake Container Services, Databricks, or your own GPU cluster. Behavioural data never leaves your environment; no shared model corpus across customers. ### Cross-Domain Transfer Behavioural knowledge learned in one domain transfers to another. Purchase patterns improve fraud detection; engagement signals sharpen churn predictions — cross-domain transfer that no single-purpose model can replicate. ## No more plumbing pipelines BaseModel.ai replaces the entire traditional ML stack — from feature engineering to model serving — with a single foundation model. - **10/12** — RelBench tasks won — BaseModel.ai is the new state of the art on 10 of the 12 standard RelBench tasks; the remaining two stay within the leading cluster. - **12 h** — Foundation training — ~8 B events / ~18 M clients on 1× NVIDIA A100. RelBench tables train in 15–30 min on 1× H100. - **9 ms** — 6 M-item recs / user — Full ranking over a 6 M-item catalog on 1× H100. Sub-millisecond for classification heads. - **Zero** — Feature engineering — Automatic feature creation from raw behavioural events via Cleora-NX + TREMDE (Synerise's proprietary internal extensions of the published Cleora and EMDE algorithms) — no manual feature work. ## Real-World Deployment Impact Across the Synerise platform powered by BaseModel — numbers from 340+ production deployments. - **340+** — Production Deployments — Active enterprise deployments across 4 continents - **2.8 B** — Daily Predictions — Real-time predictions served every day - **+23 %** — Avg. Revenue Lift — Average incremental revenue for retail customers - **+8.2 pp** — Model Accuracy Gain — Average AUC improvement over incumbent ML models - **14 days** — Deployment Speed — Median time from kickoff to production predictions - **67 %** — Cost Reduction — Reduction in ML infrastructure and team costs ## Trusted by industry leaders Companies across industries rely on BaseModel.ai to power their behavioral AI. > "BaseModel.ai allows us to personalize video recommendations at a scale and accuracy we couldn't achieve with our previous ML stack." — ML Engineering Team, Dailymotion > "Embedding generation is superfast — under 5 minutes compared to approximately 20 hours with GraphSAGE. The performance improvement is transformative." — Data Science Team, Zomato > "BaseModel.ai is the core engine for all of our AI services. It powers personalization, predictions, and recommendations across every customer touchpoint." — AI Platform Team, Synerise > "We can now tailor communication to each customer's individual preference and predicted behavior — something that was impossible before." — Marketing Technology Team, Modivo ## Limitless predictions. One model. Any industry. Select an industry to see what BaseModel.ai can answer — out of the box. Industries: General, Retail, Travel, Customer Service, Telco, Health, Insurance, Gaming, Banking, Ecommerce, Fashion, Automotive, Software, Payments, Security, News & Publishing - **General** — "How do daily customer interactions influence their future behaviors?" (360° behavioral understanding, Zero-shot ready) - **Retail** — "Which customers will churn in the next 30 days and what offers will retain them?" (Churn prediction, Offer optimization) - **Travel** — "Which travelers are most likely to upgrade and when should we reach them?" (Upsell prediction, Timing optimization) - **Customer Service** — "Which support tickets will escalate and how to prevent it?" (Escalation prediction, Proactive routing) - **Telco** — "Which subscribers are at risk and what intervention prevents churn?" (Subscriber retention, Proactive engagement) - **Health** — "How can patient engagement patterns predict treatment adherence?" (Adherence prediction, Patient engagement) - **Insurance** — "Which policyholders are most likely to file a claim next quarter?" (Claims prediction, Risk scoring) - **Gaming** — "Which players are about to stop playing and what keeps them engaged?" (Player retention, Engagement optimization) - **Banking** — "Which transaction patterns indicate potential fraud in real time?" (Fraud detection, Real-time scoring) - **Ecommerce** — "What product will each customer purchase next — and when?" (Next-best-offer, Purchase prediction) - **Fashion** — "Which styles will each customer prefer next season?" (Trend prediction, Personal styling) - **Automotive** — "When will each customer be ready for their next vehicle purchase?" (Purchase timing, Lead scoring) - **Software** — "Which trial users will convert to paid and what features drive conversion?" (Conversion prediction, Feature attribution) - **Payments** — "Which transactions are fraudulent vs. legitimate in real time?" (Fraud detection, Transaction scoring) - **Security** — "What behavioral anomalies indicate a potential security breach?" (Anomaly detection, Threat scoring) - **News & Publishing** — "Which content will maximize engagement for each reader segment?" (Content optimization, Audience segmentation) ## One model, nine superpowers BaseModel.ai handles every major behavioral prediction task — out of the box. ### Propensity prediction Predict which customers are most likely to convert, purchase, or take a desired action. ### Churn prediction Identify at-risk customers before they leave and trigger proactive retention campaigns. ### Recommendations Deliver hyper-personalized product, content, and next-best-action recommendations. ### Anomaly detection Detect unusual behavioral patterns in real time — from fraud to system anomalies. ### Profile scoring Score every customer on any dimension — engagement, lifetime value, risk, satisfaction. ### Profile matching Find lookalike audiences and similar customer profiles across your entire base. ### Hyper-segmentation Go beyond basic demographics — segment by behavioral patterns, intent, and predicted outcomes. ### Time to event predictions Predict when events will happen — next purchase, next visit, next churn event. ### Uplift estimation Measure the incremental impact of actions — which customers will respond to treatment. ## Model Governance & Explainability Enterprise-grade governance built into every layer — from training data lineage to production prediction auditing. ### YAML-Defined Targets Every fine-tuned scenario is described by a single YAML config — source tables, target function, training window. Reproducible by design; rerunning the config rebuilds the same model. ### Configurable Sample Weights Per-event sample weights and target balancing controls let teams tune fairness and recency trade-offs at training time, with full visibility into how examples are weighted. ### Self-Hosted Data Sovereignty Deploy in Snowflake Container Services, Databricks, or your own GPU cluster. Behavioural data and embeddings stay inside your environment — no shared model corpus across customers. ### Reproducible Configs & Lineage Foundation training and every fine-tune run are pinned to a YAML config and a connector schema, so you can trace any prediction back to the exact tables, time window, and parameters that produced it. ### Versioned Foundations & Adapters Foundation checkpoints and per-scenario fine-tunes are versioned independently. Roll back a scenario without retraining the foundation; promote a new foundation when you're ready. ### Monitoring Hooks Streamed metrics for input distribution, prediction confidence, and business KPI correlation — wire into your existing observability stack to catch drift early. ## Simple, secure and quick deployment Self-hosted by design. BaseModel runs in your own cloud — your behavioural data never leaves your environment. Platforms: Google Cloud, AWS, Snowflake, Azure, Databricks, ClickHouse - Deploy in your own cloud — AWS, GCP, or Azure - Delivered as a Docker image for maximum portability - No data duplication — model runs where your data lives - Fully auditable — every prediction logged and traceable ## Blazingly fast. No clusters needed. Train on ~18 M clients / ~8 B events in 12 h on a single A100. Serve classification and regression at sub-millisecond latency, and 6 M-item recommendations in 9 ms per user on 1× H100. No distributed infrastructure required. ### RelBench MAP — BaseModel vs. best baseline (higher is better) - rel-amazon review — vs. ContextGNN — BaseModel 2.53 / ContextGNN 1.63 - rel-amazon rate — vs. ContextGNN — BaseModel 3.06 / ContextGNN 2.25 - rel-hm purchase — vs. ContextGNN — BaseModel 3.67 / ContextGNN 2.93 - rel-avito ad-visit — vs. RelGNN — BaseModel 4.68 / RelGNN 3.94 Source: BaseModel paper, RelBench (12 tasks). BaseModel matches or exceeds the best published baseline on 10 of 12 tasks; four largest wins shown. Sub-millisecond latency — Per-request inference time for classification and regression at production scale - **18 M+** — Clients per training run — ~8 B events on 1× A100 in 12 h - **1** — Single GPU — No cluster required - **<1 ms** — Classification latency — Sub-ms per request - **9 ms** — 6 M-item recs — Per user on 1× H100 ## Recognized by the scientific community BaseModel.ai is built on Synerise's published research — the Cleora and EMDE papers (ICONIP 2021) and the BaseModel preprint — and rolls up under Synerise's broader 40+ publications across NeurIPS, KDD, ICONIP, and ACM RecSys. ### Preprint 2025 — BaseModel: A Foundation Model for Behavioral Data The BaseModel.ai paper. Defines the foundation-model formulation for behavioural event streams and the benchmark suite reported on /research. Production BaseModel runs Cleora-NX, TREMDE, and a customized FFN backbone — proprietary internal extensions of the published Cleora and EMDE foundations, not described in the paper. ### ICONIP 2021 — Cleora: A Simple, Strong and Scalable Graph Embedding Scheme Synerise's published hypergraph-embedding scheme. Defines the deterministic, parameter-free update T_{k+1} = normalize(P·T_k). The open-source predecessor of Cleora-NX, the proprietary engine that powers BaseModel.ai in production. ### ICONIP 2021 — An Efficient Manifold Density Estimator for All Recommendation Systems The EMDE paper. Introduces compact density sketches whose sparse codes compose under summation. The open-source predecessor of TREMDE, the proprietary, temporally-aware density-sketch engine inside BaseModel.ai. ## Make a single data scientist 10x more effective. BaseModel.ai reduces the modeling lifecycle from months to days and supercharges behavioral ML at every level. - Book a Demo - Explore Platform --- # Platform — https://basemodel.ai/platform Platform ## The behavioral AI platform From raw events to production predictions in days. A single self-supervised model that replaces your entire ML stack. - **Auto** — Sketch Shape (data-tuned) - **Sub-ms** — Classification Latency - **A100** — Recommended GPU (CUDA 12+) - **7** — Native Connectors ## Architecture Five layers from raw data to governed predictions — each designed for enterprise scale and reliability. ### Layer 1 — Data Ingestion Connect a warehouse, lake, or Parquet table and declare your event tables in YAML. The pipeline normalizes timestamps, joins related tables, and prepares events for sketching — no manual feature engineering. - 7 native warehouse / lakehouse connectors - YAML-declared schemas - Joined event + attribute tables - Temporal entity splits ### Layer 2 — Cleora-NX — Proprietary Hypergraph Embeddings Cleora-NX — Synerise's proprietary graph-embedding engine, building on the research published at ICONIP 2021. Production-only; with substantial performance and functionality extensions — not open-sourced. Deterministic (no training variance), parameter-free, runs in time proportional to the number of hyperedges, and scales to graphs with millions of nodes and billions of interactions. CPU-only. - Cleora-NX — proprietary, never published; the open-source Cleora is its predecessor - Multi-modality + temporal-interaction extensions on top of the public Cleora update - Deterministic, parameter-free, CPU-only - Scales to billions of interactions ### Layer 3 — TREMDE — Proprietary Density Sketches Cleora-NX embeddings (plus any text, image, or tabular embeddings) are aggregated into fixed-size density sketches by TREMDE — Synerise's proprietary, temporally-aware extension of the open-source EMDE algorithm (Synerise, ICONIP 2021). TREMDE produces sparse codes that are composable under summation — adding two entities' codes yields the code for their combined behaviour — and adds temporal-interaction modelling, modality-specific extensions, and other internal improvements not described in the original EMDE paper. The sketch shape itself is auto-tuned by the pipeline from your data — no manual config, no hyperparameter sweep. Each modality is sketched independently and concatenated; profiles are constructed on-the-fly with cost linear in interactions; new items with content features get meaningful sparse codes immediately (true cold-start support, no retraining). - TREMDE — proprietary, never published; the open-source EMDE is its predecessor - Adds temporal-interaction modelling on top of public EMDE - Sketch shape auto-tuned from your data — zero manual config - Cold-start support out of the box ### Layer 4 — Customized FFN Backbone Not a textbook MLP. A customized feed-forward backbone purpose-built for behavioural sequences, carrying proprietary inductive biases tuned to the structure of event data. No attention, no recurrence, no sequence operations — yet it captures temporal structure, cross-feature interactions, and modality fusion that an off-the-shelf MLP cannot. Training objective is distributional matching: cross-entropy between predicted and true depth distributions across modalities. - Captures temporal structure without recurrence or attention - Cross-feature interactions and modality fusion in one pass - Single foundation, many scenarios — no per-task retraining - Distributional matching loss; sub-millisecond inference at production scale ### Layer 5 — Inference & Serving Candidate items are encoded into TREMDE sparse codes (composable under summation). Predictions are scored using a proprietary aggregation method optimized for behavioral sketches. Sub-millisecond classification / regression; recommendation latency in single-digit milliseconds at scale. - Sub-ms classification / regression - 9 ms per user on 6M-item catalog (1× H100) - Proprietary aggregation tuned for behavioural sketches - Self-hosted by design ## Core Capabilities ### Self-Supervised Learning The foundation model is trained on raw behavioral sketches without labels. It learns to predict future user behavior from past behavior — no task-specific labels required. ### Behavioral Sketches Each user, product, and event modality is encoded into a fixed-size density sketch — shape auto-tuned by the pipeline from your data. Similar behavior produces similar sketches, and new items with content features get meaningful codes immediately. ### Multi-Task Foundation One foundation model powers many scenarios — churn, LTV, next-best-action, recommendations, regression — by training a lightweight scenario head on top of the same sketches. ### Production Serving Sub-millisecond inference for classification and regression. Recommendation scoring runs in single-digit milliseconds per user, even on multi-million-item catalogs. ### Self-Hosted by Design BaseModel deploys inside Snowflake Container Services, Databricks, or your own GPU cluster. Behavioral data never leaves your environment, and there is no shared model corpus across customers. ### Iterative Retraining Re-run the YAML-driven pipeline on a schedule to refresh sketches and the foundation backbone as new behavioral data arrives — no manual feature pipelines to maintain. ## Cleora-NX · Proprietary ### Cleora-NX — the engine inside BaseModel BaseModel ships with 5-years of proprietary, in-house improvements over prior open-source publications. Cleora-NX is Synerise's proprietary graph-embedding engine, building on the research published at ICONIP 2021. Production-only, with substantial performance and functionality extensions — not open-sourced. Cleora-NX is the hypergraph-embedding stage of the BaseModel.ai pipeline. It runs in time proportional to the number of hyperedges, scales to graphs with millions of nodes and billions of interactions, and converges in a handful of iterations on large graphs in minutes — on CPU. - Read the (public) Cleora paper - Cleora (OSS predecessor) Heads-up: the public Cleora repo and the ICONIP 2021 paper describe the foundations. They do not describe Cleora-NX. Cleora-NX is Synerise IP and is not open-sourced. - **Proprietary** — Cleora-NX is internal Synerise IP - **Deterministic** — No random walks, no negative sampling - **Multi-modal** — Graph + text + image + tabular - **Temporal** — Handles temporal interactions natively ## Integrations Connect to any data source. BaseModel.ai integrates with your existing infrastructure. - Snowflake - BigQuery - Azure Synapse - Databricks - Hive - ClickHouse - Parquet ## Governance & Compliance ### Self-Hosted Deployment Runs inside Snowflake Container Services, Databricks, or your own GPU cluster. Behavioral data never leaves your environment. ### No Shared Corpus Each customer's foundation model is trained on their own data. There is no cross-customer model corpus. ### Compatible With Your Posture Self-hosted by design — supports your existing SOC 2, HIPAA, GDPR, and EU AI Act posture rather than imposing its own. ### Reproducible Pipelines Cleora-NX is deterministic. Same data + same YAML produces the same sketches and embeddings, run after run. ### Versioned Artifacts Foundation models, sketches, and scenario heads are stored as versioned artifacts in your warehouse or object store. ### Logged Training Runs Training runs log to Weights & Biases or CSV. Inspect metrics, compare runs, and audit the pipeline end-to-end. ## Ready to see it in action? Book a demo to see how BaseModel.ai can replace your ML stack with a single foundation model. Book a Demo --- # How it works — https://basemodel.ai/how-it-works How It Works ## From raw data to predictions BaseModel.ai uses a two-stage approach: build a foundation model from your behavioral data, then train lightweight scenario heads that reuse the foundation embeddings — no manual feature engineering required. ## The Two-Stage Flow First, build a foundation model that learns universal behavioral patterns. Then, train scenario heads for specific business objectives that reuse the foundation embeddings. ### Stage 1 — Foundation Model Cleora-NX (Synerise's proprietary, internal extension of the open-source Cleora algorithm published at ICONIP 2021) turns your event log into a hypergraph and runs deterministic power iterations to produce node embeddings. TREMDE (Synerise's proprietary, temporally-aware extension of the open-source EMDE algorithm, also ICONIP 2021) then aggregates those embeddings (along with any text, image, or tabular features) into per-user density sketches. A customized FFN backbone — purpose-built for behavioural sequences, with proprietary inductive biases tuned to event data — learns to predict each user's future sketch from their past sketch. BaseModel ships with 5-years of proprietary, in-house improvements over prior open-source publications. - Cleora-NX hypergraph + power iteration (proprietary) - TREMDE density sketches (proprietary; shape auto-tuned from your data) - Customized FFN backbone — purpose-built for behavioural sequences; no transformers, no attention, no recurrence - Self-supervised — no labels required ### Stage 2 — Scenario Models Lightweight scenario heads attach to the foundation and reuse the same sketches. Each head targets one of the five supported task types — binary, multiclass, multilabel, regression, or recommendation — and trains on your labeled examples in a fraction of the time required for a from-scratch model. - Reuses foundation sketches - 5 task types supported - Minimal labeled data needed - Rapid iteration cycles ## Visual Pipeline Five steps from raw behavioral data to production-ready predictions. ### Step 1 — Data Connection & YAML Connect a warehouse, lake, or Parquet table — Snowflake, BigQuery, Azure Synapse, Databricks, Hive, ClickHouse, or Parquet. Declare your event tables, joined attribute tables, main entity column, and date column in YAML. ### Step 2 — Cleora-NX Hypergraph Embedding Each behavioral event is treated as a hyperedge linking all participating entities. Cleora-NX — Synerise's proprietary, internal extension of the open-source Cleora algorithm (ICONIP 2021), with multi-modality and temporal-interaction extensions not in the published version — expands these into pairwise relations, builds a row-normalized transition matrix, and runs power iterations on random unit vectors. Deterministic, CPU-only, parameter-free. ### Step 3 — TREMDE Density Sketches Cleora-NX embeddings (plus any text, image, or tabular features) are aggregated into per-user density sketches by TREMDE — Synerise's proprietary, temporally-aware extension of the open-source EMDE algorithm (ICONIP 2021), with temporal-interaction modelling and modality-specific extensions not in the published version. Sketch shape is auto-tuned by the pipeline from your data — no manual config required. ### Step 4 — Customized FFN Backbone Training A customized feed-forward backbone — purpose-built for behavioural sequences, with proprietary inductive biases tuned to the structure of event data — is trained to predict the user's future sketch from the past sketch. Loss is cross-entropy between predicted and true depth distributions. No attention, no recurrence, yet it captures temporal structure and cross-feature interactions that an off-the-shelf MLP cannot. ### Step 5 — Scenario Heads & Scoring A lightweight scenario head — binary, multiclass, multilabel, regression, or recommendation — reuses the foundation sketches. Scoring is sub-ms for classification / regression; ~9 ms per user to score a 6 M-item catalog on 1× H100. ## The Adoption Path Five stages from initial setup to predictions landing in your warehouse — typically completed in well under a week on 1× A100. ### Stage 1 — Connect Data Point BaseModel at a warehouse, lake, or Parquet table and write a YAML config describing your event and attribute tables. ### Stage 2 — Cleora-NX + TREMDE The pipeline builds the hypergraph, runs Cleora-NX power iterations (Synerise's proprietary extension of the open-source Cleora), and aggregates the resulting embeddings into per-user density sketches via TREMDE (Synerise's proprietary temporally-aware extension of the open-source EMDE). ### Stage 3 — Train Foundation The customized FFN foundation backbone — purpose-built for behavioural sequences, with proprietary inductive biases tuned to event data — trains on sketches to predict future user behaviour. Reference run: ~18 M clients / ~8 B events in 12 h on 1× NVIDIA A100; smaller datasets train proportionally faster (RelBench tables in 15–30 min on 1× H100). ### Stage 4 — Train Scenario Head A lightweight scenario head — churn, LTV, recommendation, regression — is trained on top of the foundation, reusing the sketches. ### Stage 5 — Score & Deliver Predictions are written back to your warehouse or to TSV files. Your existing CDP, CRM, or BI stack picks them up via native warehouse connectors. ## Automated Feature Engineering No manual pipelines. The foundation model automatically engineers features from raw behavioral data. ### Traditional Approach - Months of manual feature engineering - Domain experts required for each data type - Brittle pipelines that break on schema changes - Features must be redesigned per use case ### BaseModel.ai Approach - Automated feature engineering from raw events - Self-supervised learning — no labels needed - Schema-adaptive — handles data changes gracefully - Universal embeddings work across all use cases ## Key Benefits ### Predictive Power On the public RelBench suite, BaseModel matches or exceeds the best baseline on 10 of 12 tasks, with double-digit MAP gains over ContextGNN and RelGNN on several recommendation tables. ### Workflow Efficiency Skip months of hand-crafted feature engineering. The YAML-driven Cleora-NX + TREMDE pipeline turns raw event tables into model-ready sketches automatically. ### Comprehensive Data Utilization Every interaction, every event type, and any text / image / tabular features are sketched as independent modalities and concatenated. No signal is wasted. ### Quick Time-to-Value A typical proof-of-concept goes from YAML to scored predictions in well under a week on 1× A100, depending on data volume. ### Reproducibility Cleora-NX is deterministic — same data + same YAML produces the same embeddings. Training runs log to Weights & Biases or CSV for audit and comparison. ### Single Foundation, Many Scenarios Train one foundation backbone, then attach lightweight scenario heads — churn, LTV, recommendation, regression — that all reuse the same sketches. ## Ready to get started? See how BaseModel.ai turns your behavioral data into warehouse-resident predictions, end-to-end. Book a Demo --- # Data models — https://basemodel.ai/data-models Data & Models ## Data model & training Understand the data model, requirements, and training modes that power BaseModel.ai's behavioral predictions. ## Data Model BaseModel.ai's data model is built around entities, events, and interactions. ### Entities Core objects in your data model — customers, products, sessions, devices. Each entity has a unique identifier and associated behavioral events. - Unique entity identifiers - Multi-type support - Profile attributes - Hierarchical relationships ### Main Entity The primary entity around which the model is built — typically customers or users. All predictions and embeddings are generated for the main entity. - Central to the model - Embedding target - Prediction subject - Typically users or customers ### Events Timestamped behavioral interactions — page views, purchases, clicks, logins. Events form the sequential data that the foundation model learns from. - Timestamped actions - Sequential ordering - Multi-type events - Feature-rich payloads ### Interactions Relationships between entities — customer-product interactions, user-content engagements. These form the graph structure used for embedding generation. - Entity relationships - Graph structure - Temporal dynamics - Weighted connections ## Minimum Data Requirements The minimum data volumes needed for effective model training. - **1+** — Event tables required. BaseModel.ai needs at least one event table. Attribute tables are welcome and joined automatically by entity ID — no manual feature engineering. - **≥ 10,000** — Main-entity profiles. Minimum number of unique main-entity profiles (e.g. customers, users) needed for meaningful foundation training. - **≥ 100,000** — Interactions / month. Minimum monthly interaction volume so the foundation can capture stable behavioral patterns. - **≥ 3 months** — Frequent interactions. Minimum history depth when interactions are frequent (e.g. e-commerce sessions, ad clicks, content views). - **≥ 1 year** — Infrequent interactions. Minimum history depth when interactions are infrequent (e.g. insurance claims, big-ticket purchases) so seasonality is observable. ## Hardware Requirements ### GPU - Recommended: NVIDIA A100 or better - Minimum: multi-GPU cluster of A10s / L40s - CUDA 12+ ### Memory - 240 GB RAM - 32 CPU cores - Scales with dataset size ### Storage & Runtime - 1 TB disk - Docker-capable environment - Training and inference scale linearly with GPU count ### Production performance profile From the BaseModel.ai docs Reference deployment from the canonical hardware appendix — ~8 B events / ~18 M unique clients / ~1 M products on a single NVIDIA A100. - **12 h** — Foundation training · 1× NVIDIA A100 - **10 h** — Scenario fine-tuning · 16,000 brands · 1× A100 - **2,718** — Inference clients/sec/GPU · Full ranking throughput - **~8 B** — Events trained · ~18 M clients · ~1 M products ## Training Options Two stages: pre-train the foundation once with 01_pretrain.py, then attach as many scenario heads as you need with the per-task scripts. ### Stage 1 — Foundation Pre-training (01_pretrain.py) Run 01_pretrain.py once per data slice. It executes Cleora-NX hypergraph embedding (Synerise's proprietary internal extension of open-source Cleora), TREMDE density sketching (Synerise's proprietary temporally-aware extension of open-source EMDE), and trains the customized FFN foundation backbone (purpose-built for behavioural sequences, with proprietary inductive biases tuned to event data) against the distributional matching objective. The output is a pinned foundation checkpoint and the encoding state used to produce TREMDE sparse codes for new entities. - Single Python entry point: 01_pretrain.py - Cleora-NX + TREMDE + customized FFN foundation backbone (proprietary) - Outputs reusable foundation checkpoint - Reused by every downstream scenario head ### Stage 2 — Per-Scenario Training (01_train.py / 02_test.py / 03_predict.py) Define a prediction task, train a scenario head, evaluate, and score — all with simple Python scripts that ship with the product. Task types: binary classification, multiclass, multilabel, regression, and recommendation. - 01_train.py — fits the scenario head on frozen sketches - 02_test.py — held-out evaluation - 03_predict.py — writes predictions to the warehouse - Many scenario heads share one foundation ## Per-Entity Split Points With `split.type: entity`, BaseModel.ai picks an independent split point in time for every main-entity profile inside the train/validation window bounded by `training_validation_end`. Events before each entity's split point feed Cleora-NX, TREMDE, and the foundation backbone; events after it become that entity's per-entity validation target. The test window is a separate, explicit date range — the model never sees the future of any entity during training. ### Training — Per-entity past For each main entity, all events before its individually chosen split point feed Cleora-NX, TREMDE, and the foundation backbone. ### Validation — Per-entity future Events after each entity's split point — bounded by training_validation_end — form that entity's validation target. The training : validation ratio (e.g. 90 / 10) controls where each split point lands. ### Test — Explicit window An explicit test window — split.test.start_date → end_date in YAML — held out for the final scenario-head evaluation. No entity has any events from this window during training. Earliest data — Most recent ## Foundation vs Scenario Models ### Foundation Model The universal behavioural model trained via self-supervised learning. Cleora-NX (Synerise's proprietary extension of open-source Cleora) produces deterministic hypergraph embeddings, TREMDE (Synerise's proprietary temporally-aware extension of open-source EMDE) turns them into per-user density sketches with shape auto-tuned by the pipeline from your data, and a customized FFN backbone — purpose-built for behavioural sequences, with proprietary inductive biases tuned to event data — learns to predict each user's future sketch from the past. - Self-supervised pre-training - No labels required - Cleora-NX + TREMDE + customized FFN backbone (proprietary) - Sketch shape auto-tuned from your data - Trained once, used many times ### Scenario Models Task-specific models that use foundation embeddings as input. Each scenario model targets a specific business objective — churn prediction, recommendations, LTV — with minimal additional training. - Reuses foundation embeddings - Task-specific objectives - Minimal labeled data needed - Fast training iterations - Multiple scenario heads per foundation ## Ready to train your model? Get started with BaseModel.ai and train your first behavioral foundation model. - Book a Demo --- # Benchmarks — https://basemodel.ai/benchmarks Benchmarks ## BaseModel.ai vs the field, in numbers Independent and in-house benchmarks across sequential recommendation (Amazon Beauty / Sports / Toys, Amazon Books) and relational deep learning (RelBench), plus a live production A/B test. Higher AUROC, MAP, NDCG, and Recall are better; lower MAE is better. ### Headline results from the BaseModel paper. Headline results — From the BaseModel paper Across sparse Amazon categories, large-scale Amazon Books, MovieLens, and the multi-table RelBench suite, a single pretrained BaseModel.ai backbone is reported to beat Meta's HSTU, Google's TIGER, beeFormer, ContextGNN, and RelGNN under each method's own published protocol. All numbers below are taken directly from the BaseModel preprint. - **+138%** — NDCG@10 vs HSTU · Amazon Books · Meta's flagship sequential model - **+82%** — NDCG@5 vs TIGER · Amazon Sports · Google's NeurIPS '23 generative SOTA - **10 / 12** — RelBench tasks: SOTA · Beating ContextGNN and RelGNN on their own benchmark - **+21%** — Open rate, in production · Live A/B test, ~2M emails, major EU retailer ### BaseModel vs Meta's HSTU (Amazon Books) Sparser, longer-tail benchmark. 674K items, 14.5 interactions per user. HSTU-large is Meta's biggest published configuration. | Metric | HSTU-lg | BaseModel | Uplift | | --- | --- | --- | --- | | HR@10 | 4.69 | 9.52 | +103% | | NDCG@10 | 2.57 | 6.12 | +138% | | NDCG@200 | 5.08 | 8.39 | +65% | Source: BaseModel paper, Table 3. HSTU-large baseline; BaseModel mean ± std over 5 seeds. All values × 100. ### Outpacing Google's TIGER, by 54–82 % (Sparse Amazon) Sparse Amazon categories, fewer than 9 interactions per user. TIGER (NeurIPS 2023) was the prior generative-retrieval state of the art. | Metric | TIGER | BaseModel | Uplift | | --- | --- | --- | --- | | Beauty · NDCG@10 | 3.84 | 6.80 | +77% | | Sports · NDCG@5 | 1.81 | 3.30 | +82% | | Toys · NDCG@10 | 4.32 | 6.82 | +58% | Source: BaseModel paper, Table 2. TIGER baseline from Rajput et al., NeurIPS 2023. All values × 100. ### Independent reproduction vs LIGER Synerise SAIR · Independent reproduction LIGER (Nov 2024) was the new sequential-recommendation state of the art and itself a step beyond TIGER. The Synerise SAIR research team reproduced the LIGER protocol verbatim and ran BaseModel.ai head-to-head. The cold-start gaps below stack on top of LIGER's own gains over TIGER. - **+263%** — NDCG@10 · Sports · cold-start - **+243%** — Recall@10 · Sports · cold-start - **+103%** — NDCG@10 · Toys · cold-start - **+24–60% NDCG@10** warm-start, all 3 datasets - **Every metric, every dataset** — no exceptions Source: Synerise SAIR — *Beyond LIGER*. Cold-start range across both BaseModel data-prep configurations reported by SAIR; both beat LIGER on every cell. ### RelBench — relational deep-learning suite Matches or beats best baseline on 10 of 12 tasks | RelBench task | Metric | BaseModel.ai | Best prior | Δ | | --- | --- | --- | --- | --- | | rel-amazon · review (recs) | MAP×100 | 2.53 | 1.63 (ContextGNN) | +55% | | rel-amazon · rate (recs) | MAP×100 | 3.06 | 2.25 (ContextGNN) | +36% | | rel-hm · purchase (recs) | MAP×100 | 3.67 | 2.93 (ContextGNN) | +25% | | rel-avito · ad-visit (recs) | MAP×100 | 4.68 | 3.94 (RelGNN) | +19% | | rel-avito · clicks (clf) | AUROC | 68.54 | 68.23 (RelGNN) | +0.31 | | rel-hm · churn (clf) | AUROC | 71.25 | 70.93 (RelGNN) | +0.32 | | rel-avito · visits (clf) | AUROC | 66.25 | 66.18 (RelGNN) | +0.07 | | rel-amazon · churn (clf) | AUROC | 71.02 | 70.99 (RelGNN) | +0.03 | | rel-amazon · user-ltv (reg) | MAE (lower) | 14.26 | 14.23 (RelGNN) | ≈ tied | | rel-amazon · purchase (recs) | MAP×100 | 2.89 | 2.93 (ContextGNN) | −1% (within variance) | | rel-stack · engage (clf) | AUROC | 90.00 | 90.75 (RelGNN) | −0.75 | | rel-stack · comment (recs) | MAP×100 | 13.06 | 14.00 (RelGNN) | −7% | All 12 tasks from the BaseModel paper, Table 5 — six recommendation, five classification, one regression. BaseModel matches or beats the best graph-neural-network baseline (RelGNN, ContextGNN) on 10. The two exceptions both involve rel-stack, where tight Stack-Exchange community structure favors local message passing. Each foundation model trains on 1× NVIDIA H100 in 15–30 minutes per RelBench table. ### Cold-start: 71 % of items unseen at training time (rel-avito · ad-visit) In rel-avito, 71 % of test ads were never seen during training — the scenario where every classical recommender goes blind. BaseModel.ai's content-aware density sketches let unseen items participate in recommendations from the second they appear, with no retraining. Adding metadata to the profile turns the model from unusable to dominant on the same data. - **+31.8 %** — mAP · Interactions only → +metadata - **+65.7 %** — Precision · On the same cold-start split - **+30.3 %** — Recall · Cold items recovered, not buried - **80.7 %** — Cold-start surface rate · Of top-K recs are new items Source: BaseModel paper, Table 6. New items receive meaningful representations from content features alone — no retraining, no warm-up. ### Production A/B test, ~2 M emails (Major EU retailer) Live deployment in the email-marketing channel of a large European omnichannel retailer. BaseModel.ai replaced an incumbent collaborative-filtering + XGBoost ranker built on hand-engineered features. Random assignment to control or treatment across multiple weekly campaigns and seasonal periods. - **+21 %** — Open rate - **+18 %** — Click-through - **+7 %** — Conversion Source: BaseModel paper, §6 — Deployment Experience. Lift over the incumbent personalization stack across approximately 2 million emails. ### Pretraining = ~2× faster convergence (Foundation transfer) A backbone initialized from BaseModel.ai's self-supervised foundation reaches the same Recall@25 in roughly half the training time as the same architecture from scratch — and finishes ahead. The same lift appears on a churn classification task that the foundation never saw, evidence that pretraining captures generic behavioral structure rather than task-specific patterns. - **~2×** — Faster to SOTA - **15–30 min** — Per RelBench task Source: BaseModel paper, §5.3 — Effect of Foundation Pretraining (Figures 3 & 4). Single NVIDIA H100, batch 512. --- # Use cases — https://basemodel.ai/use-cases Use Cases ## What you can predict From churn prediction to personalized recommendations — BaseModel.ai supports a wide range of prediction tasks through its foundation model approach. ## Prediction Types The foundation model supports multiple prediction task types, each powered by the same behavioral embeddings. ### Binary Classification **Churn Prediction** Predict binary outcomes — will a customer churn or not? Will they convert or not? Binary classification is the most common use case for behavioral AI. **Example:** Predict which customers will churn in the next 30 days. The model analyzes engagement patterns, purchase frequency, and behavioral sketches to identify at-risk customers before they leave. - AUC-ROC scoring - Precision/Recall tradeoff - Score calibration - Threshold optimization ### Multiclass Classification **Favorite Brand Campaigns** Classify entities into multiple categories. Identify which campaign, brand, or segment a customer belongs to based on their behavioral patterns. **Example:** Predict which brand campaign each customer is most likely to engage with. The model learns preferences from historical interactions and behavioral similarity to other customers. - Top-K accuracy - Per-class precision - Confusion matrix analysis - Macro/micro F1 scores ### Multi-Label Classification **Multiple Interests** Assign multiple labels simultaneously. A customer can belong to multiple segments, have multiple interests, or qualify for multiple offers at the same time. **Example:** Identify all product categories a customer is interested in. Unlike multiclass, each customer can have multiple active interests — fashion AND electronics AND home goods. - Per-label AUC - Macro / micro F1 - Subset accuracy - Label correlation analysis ### Regression **LTV & Demand Forecasting** Predict numeric values — customer lifetime value, expected basket size, days-to-next-purchase, demand forecasts. Regression heads attach to the same foundation sketches as the classification heads. **Example:** Forecast 90-day customer LTV from behavioral sketches. The foundation captures purchase rhythm, engagement depth, and category affinity — signals that traditional LTV models built on aggregated features miss. - MAE / MAPE - RMSE - R² coefficient - Quantile calibration ### Recommendations **Acquisition & Retention** Generate personalized product, content, or action recommendations. Support both acquisition (new customer) and retention (existing customer) scenarios. **Example:** Recommend products for each customer based on their behavioral sketches. The foundation captures taste and intent from behavioral signals — not just purchase history. Sub-10 ms latency per user even on multi-million-item catalogs. - MAP@K - NDCG@K - Hit rate @ K - Coverage & diversity ## Prediction & Scoring Flow From defining your target to deploying predictions — a streamlined flow. 1. **Define Target Function** — Write a Python target function that says what to predict — a binary label, a category, a numeric value, or a ranked candidate set. 2. **Train Scenario Head** — A lightweight scenario head trains on top of the foundation sketches with your labels — fast because the heavy representation work is already done. 3. **Generate Predictions** — Score your entire entity base. Each entity receives a numeric score plus the metric appropriate to its task type. 4. **Evaluate & Deploy** — Review metrics, compare against baselines, and write predictions back to your warehouse or to TSV files for downstream systems to pick up. ## Case Study Reference ### H&M Kaggle Dataset The BaseModel.ai documentation includes an end-to-end tutorial using the H&M Kaggle dataset — demonstrating the complete pipeline from data preparation to trained recommendation models. - Real-world fashion recommendation dataset - 31M+ customer transactions - Complete pipeline walkthrough - Foundation model + recommendation scenario View Tutorial - **31M+** — Transactions - **1.3M** — Customers - **105K** — Products - **MAP@12** — Ranking Metric ## Production Case Study From the BaseModel research paper: a large European omnichannel retailer ran an email A/B test against an incumbent collaborative-filtering + XGBoost ranker. - **+21%** — Open Rate — Lift over the incumbent CF + XGBoost ranker. - **+18%** — Click-Through Rate — Subject lines and content selected by BaseModel scenario heads. - **+7%** — Conversion — Downstream revenue lift from the same campaign cycles. Approximately 2 million emails over multiple weekly campaign cycles. Source: BaseModel research paper (under double-blind review). ## Acquisition vs Retention Recommendations serve different objectives depending on the customer lifecycle stage. ### Acquisition Recommendations For new or early-stage customers with limited behavioural data. TREMDE's content-aware sparse codes (Synerise's proprietary, temporally-aware extension of open-source EMDE) give every new item meaningful sketch coordinates immediately — true cold-start support, no retraining required. - Cold-start capable - Inductive embeddings - Behavioral similarity to existing users - Progressive personalization ### Retention Recommendations For established customers with rich behavioral history. Leverage deep temporal patterns and purchase cycles to recommend products that maximize engagement and lifetime value. - Rich behavioral context - Temporal pattern analysis - Cross-sell & upsell - Lifetime value optimization ## Explore your use case Every business has unique prediction needs. Let us show you how BaseModel.ai applies to your specific use case. - Book a Demo - View Industries --- # Industries — https://basemodel.ai/industries Industries ## One model. Every industry. BaseModel.ai serves 14 industries with a single behavioral foundation model. No custom training per vertical — cross-domain transfer is built in. ### All Industries - **Retail** — BaseModel.ai transforms raw retail behavioral data into real-time predictions — from next-best-offer to churn prevention — without manual feature engineering. — **+23%** — Average Revenue Lift - **Banking** — A single behavioral foundation model that learns from every transaction, login, and interaction — serving fraud detection, credit scoring, and personalization simultaneously. — **99.7%** — Fraud Detection Rate - **Ecommerce** — From product discovery to post-purchase — a single model that understands every customer's intent, preferences, and purchase timing. — **+31%** — Revenue Per Visitor - **Telco** — Understand every subscriber's behavioral patterns — from usage to support interactions — to predict churn, optimize offers, and personalize engagement. — **-42%** — Churn Reduction - **Travel** — From search to booking to post-trip — predict travel intent, optimize pricing, and personalize every touchpoint. — **+26%** — Booking Conversion - **Healthcare** — Transform patient behavioral and clinical data into predictions that improve outcomes, reduce costs, and personalize care. — **+29%** — Adherence Improvement - **Insurance** — Behavioral intelligence for claims prediction, fraud detection, pricing optimization, and policyholder retention. — **+19%** — Loss Ratio Improvement - **Gaming** — Predict player behavior — from session length to purchase intent to churn risk — using a single behavioral foundation model. — **+38%** — Player Retention - **Automotive** — From online research to showroom visit to service appointment — understand every customer's automotive journey. — **+33%** — Lead Conversion - **Software / SaaS** — Predict trial-to-paid conversion, expansion opportunities, and churn risk from product usage behavioral patterns. — **+41%** — Trial Conversion - **Payments** — Real-time behavioral scoring for transaction authorization, fraud prevention, and merchant risk management. — **99.9%** — Authorization Accuracy - **Cybersecurity** — Transform security logs and user behavior into real-time threat detection — catching attacks that signature-based systems miss. — **97%** — Threat Detection Rate - **Fashion** — Understand every customer's style preferences, purchase patterns, and trend sensitivity to drive personalized fashion experiences. — **+35%** — Full-Price Sell-Through - **Customer Service** — Predict escalations, optimize routing, and personalize support — transforming reactive service into proactive engagement. — **-45%** — Escalation Rate - **News & Publishing** — Predict content preferences, optimize paywalls, and personalize the reading experience at scale. — **+52%** — Subscription Conversion --- ## Retail — https://basemodel.ai/industries/retail All Industries Retail ### Predict what every customer will buy next BaseModel.ai transforms raw retail behavioral data into real-time predictions — from next-best-offer to churn prevention — without manual feature engineering. Book a Demo **+23%** — Average Revenue Lift ### Overview Retail teams spend months building individual models for each behavioral prediction. BaseModel.ai replaces the entire ML stack with a single foundation model that learns from every customer interaction — browsing, purchasing, returns, support tickets — and predicts any individual behavior in real time. ### Use Cases - **Next-Best-Offer** — Predict the optimal product, discount, and channel for each customer at every touchpoint. — **+18%** Conversion Lift - **Churn Prevention** — Identify at-risk customers 30–90 days before churn and trigger personalized retention campaigns. — **-34%** Churn Reduction - **Customer Lifetime Value** — Predict long-term value from the first interaction using behavioral embeddings. — **92%** LTV Accuracy - **Dynamic Pricing** — Optimize pricing in real time based on demand signals, competitive landscape, and individual willingness to pay. — **+12%** Margin Improvement ### Capabilities - Unified customer profile across all touchpoints - Real-time product recommendations at sub-ms latency - Cross-channel attribution and journey analysis - Seasonal pattern detection and demand forecasting - Automatic segment discovery and micro-targeting - Privacy-preserving behavioral embeddings ### Results - **+23%** — Revenue Lift — Average incremental revenue across retail deployments - **14 days** — Time to Value — From data connection to production predictions - **85+** — Retail Clients — Enterprise retailers across 4 continents - **Sub-ms** — Latency — Real-time inference at checkout and browsing ### Questions you can answer Zero-shot — no custom model training required. - "Which customers will churn in the next 30 days?" - "What product should we recommend to each customer right now?" - "Which promotions will maximize revenue without cannibalizing margin?" - "How do browsing patterns predict purchase intent?" - "Which customer segments are emerging that we haven't identified?" ### Case Study **Top-5 European Retailer** - **Challenge** — 15-person data science team spending 6 months per model with 12 separate ML pipelines for different business questions. - **Solution** — Replaced all 12 models with a single BaseModel.ai instance trained on 2 years of behavioral data across 50M customers. - **Result** — +23% revenue lift, 67% cost reduction in ML infrastructure, deployment in 14 days. ### Ready to transform retail? See how BaseModel.ai can answer your most critical retail questions — in days, not months. Book a Demo Explore Other Industries --- ## Banking — https://basemodel.ai/industries/banking All Industries Banking ### Detect fraud and predict risk in real time A single behavioral foundation model that learns from every transaction, login, and interaction — serving fraud detection, credit scoring, and personalization simultaneously. Book a Demo **99.7%** — Fraud Detection Rate ### Overview Banking institutions struggle with siloed models for fraud, credit risk, AML, and customer engagement. BaseModel.ai unifies all behavioral signals into a single embedding space, enabling cross-task predictions from one model. ### Use Cases - **Fraud Detection** — Real-time transaction scoring using behavioral sequence patterns — catching anomalies traditional rules miss. — **99.7%** Detection Rate - **Credit Risk Scoring** — Behavioral-enriched credit models that incorporate interaction patterns beyond traditional bureau data. — **+15pp** Gini Improvement - **AML Monitoring** — Pattern detection across complex transaction networks using graph embeddings. — **-60%** False Positives - **Customer Engagement** — Predict product needs and optimal engagement timing for each customer. — **+28%** Cross-Sell Lift ### Capabilities - Sub-millisecond transaction scoring - Graph-based entity resolution across accounts - Temporal pattern detection for behavioral anomalies - Per-prediction event-level attribution for review - On-premise deployment for data sovereignty - Multi-channel behavioral fusion ### Results - **99.7%** — Fraud Detection — Detection rate with <0.1% false positive rate - **-60%** — False Positives — Reduction in AML false alerts - **45+** — Banking Clients — Tier-1 and Tier-2 banks globally - **Sub-ms** — Scoring Latency — Real-time transaction authorization ### Questions you can answer Zero-shot — no custom model training required. - "Is this transaction fraudulent based on behavioral context?" - "What is this customer's real-time credit risk?" - "Which accounts show suspicious AML patterns?" - "When is the optimal time to offer a new product?" - "Which customers are likely to increase deposits?" ### Case Study **Top-10 European Bank** - **Challenge** — Separate fraud, AML, and engagement models with 50+ ML engineers maintaining 30+ production pipelines. - **Solution** — Unified behavioral model processing 500M daily events across all customer touchpoints. - **Result** — 99.7% fraud detection, 60% fewer AML false positives, 8-week deployment. ### Ready to transform banking? See how BaseModel.ai can answer your most critical banking questions — in days, not months. Book a Demo Explore Other Industries --- ## Ecommerce — https://basemodel.ai/industries/ecommerce All Industries Ecommerce ### Personalize every moment of the shopping journey From product discovery to post-purchase — a single model that understands every customer's intent, preferences, and purchase timing. Book a Demo **+31%** — Revenue Per Visitor ### Overview Ecommerce platforms generate millions of behavioral events per hour. BaseModel.ai ingests every click, search, cart action, and purchase to build real-time customer embeddings that power personalization at scale. ### Use Cases - **Product Recommendations** — Context-aware recommendations that adapt to real-time browsing behavior, not just purchase history. — **+31%** RPV Lift - **Search Personalization** — Re-rank search results based on individual behavioral embeddings and predicted intent. — **+22%** Search CTR - **Cart Abandonment** — Predict abandonment risk and trigger real-time interventions with optimal incentives. — **-28%** Abandonment - **Demand Forecasting** — Predict product demand at SKU level using behavioral signals and temporal patterns. — **94%** Forecast Accuracy ### Capabilities - Real-time session-based personalization - Behavioral search re-ranking - Cross-device identity resolution - Visual + behavioral product similarity - Automatic A/B test optimization - Category affinity prediction ### Results - **+31%** — Revenue Per Visitor — Average lift across ecommerce deployments - **+22%** — Search CTR — Improvement in search result relevance - **60+** — Ecommerce Clients — From DTC brands to marketplaces - **200M+** — Daily Predictions — Recommendations served in real time ### Questions you can answer Zero-shot — no custom model training required. - "What product will this visitor buy in this session?" - "Which search results are most relevant for this user?" - "Will this customer abandon their cart?" - "What's the optimal discount to convert this visitor?" - "Which products will trend next week?" ### Case Study **Leading Fashion Marketplace** - **Challenge** — Generic recommendations based on collaborative filtering with 6-month model refresh cycles. - **Solution** — Real-time behavioral embeddings updating with every customer interaction across web and mobile. - **Result** — +31% revenue per visitor, +22% search CTR, deployed in 10 days. ### Ready to transform ecommerce? See how BaseModel.ai can answer your most critical ecommerce questions — in days, not months. Book a Demo Explore Other Industries --- ## Telco — https://basemodel.ai/industries/telco All Industries Telco ### Reduce churn and maximize subscriber lifetime value Understand every subscriber's behavioral patterns — from usage to support interactions — to predict churn, optimize offers, and personalize engagement. Book a Demo **-42%** — Churn Reduction ### Overview Telco operators sit on massive behavioral datasets — call records, data usage, app behavior, support tickets, billing events. BaseModel.ai transforms this complexity into actionable predictions. ### Use Cases - **Churn Prediction** — Identify at-risk subscribers 60–90 days before contract end using behavioral drift detection. — **-42%** Churn Rate - **Plan Optimization** — Recommend the optimal plan for each subscriber based on usage patterns and predicted needs. — **+18%** ARPU Lift - **Network Experience** — Predict QoS issues before they impact customer satisfaction using behavioral signals. — **-35%** Complaints - **Cross-Sell** — Identify subscribers most likely to adopt additional services — streaming, insurance, devices. — **+25%** Attach Rate ### Capabilities - CDR and behavioral data fusion - Network event correlation - Subscriber journey modeling - Real-time usage anomaly detection - Multi-SIM household resolution - Roaming behavior analysis ### Results - **-42%** — Churn Reduction — Average reduction in voluntary churn - **+18%** — ARPU Growth — Through behavioral plan optimization - **30+** — Telco Clients — Operators across 15 countries - **500M+** — Subscribers — Total subscribers modeled ### Questions you can answer Zero-shot — no custom model training required. - "Which subscribers will churn in the next 90 days?" - "What plan should we offer each subscriber at renewal?" - "Which network issues are causing the most dissatisfaction?" - "When should we reach out to prevent churn?" - "Which subscribers are ready for a device upgrade?" ### Case Study **Tier-1 European Operator** - **Challenge** — 12 separate churn models across prepaid, postpaid, and business segments with inconsistent accuracy. - **Solution** — Single foundation model trained on all subscriber behavioral data across segments. - **Result** — -42% churn, +18% ARPU, consolidated from 12 models to 1. ### Ready to transform telco? See how BaseModel.ai can answer your most critical telco questions — in days, not months. Book a Demo Explore Other Industries --- ## Travel — https://basemodel.ai/industries/travel All Industries Travel ### Anticipate every traveler's next journey From search to booking to post-trip — predict travel intent, optimize pricing, and personalize every touchpoint. Book a Demo **+26%** — Booking Conversion ### Overview Travel companies deal with complex, seasonal, and highly temporal behavioral patterns. BaseModel.ai captures the full journey context to power predictions that traditional models cannot match. ### Use Cases - **Booking Prediction** — Predict which searches will convert to bookings and optimize the path to purchase. — **+26%** Conversion Lift - **Dynamic Pricing** — Real-time pricing optimization based on demand signals, competitor data, and individual willingness to pay. — **+15%** Revenue/Seat - **Ancillary Upsell** — Predict which add-ons each traveler will purchase — upgrades, insurance, experiences. — **+34%** Ancillary Revenue - **Loyalty Optimization** — Predict member engagement and optimize reward strategies to maximize lifetime value. — **+21%** Retention Lift ### Capabilities - Temporal booking pattern analysis - Multi-leg journey prediction - Seasonal demand forecasting - Price sensitivity modeling - Cross-channel attribution - Group and family travel detection ### Results - **+26%** — Conversion — Booking conversion improvement - **+34%** — Ancillary Revenue — Through personalized upselling - **25+** — Travel Clients — Airlines, hotels, OTAs - **100M+** — Travelers — Profiles modeled annually ### Questions you can answer Zero-shot — no custom model training required. - "Will this search convert to a booking?" - "What's the optimal price for this route and traveler?" - "Which ancillary products will this traveler buy?" - "When will this loyalty member book their next trip?" - "Which destinations are trending for this segment?" ### Case Study **Major European Airline** - **Challenge** — Separate models for pricing, recommendations, and loyalty with 9-month development cycles. - **Solution** — Unified behavioral model across booking, flying, and loyalty program interactions. - **Result** — +26% booking conversion, +34% ancillary revenue, deployed in 3 weeks. ### Ready to transform travel? See how BaseModel.ai can answer your most critical travel questions — in days, not months. Book a Demo Explore Other Industries --- ## Healthcare — https://basemodel.ai/industries/health All Industries Healthcare ### Predict patient outcomes and optimize care pathways Transform patient behavioral and clinical data into predictions that improve outcomes, reduce costs, and personalize care. Book a Demo **+29%** — Adherence Improvement ### Overview Healthcare organizations generate rich behavioral data — appointment patterns, medication adherence, portal interactions, wearable data. BaseModel.ai learns from these signals to predict and prevent adverse outcomes. ### Use Cases - **Treatment Adherence** — Predict which patients are at risk of non-adherence and trigger timely interventions. — **+29%** Adherence Rate - **Readmission Prevention** — Identify patients at risk of 30-day readmission using post-discharge behavioral patterns. — **-31%** Readmissions - **Care Pathway Optimization** — Recommend optimal care pathways based on patient behavioral and clinical profiles. — **-18%** Cost per Episode - **Patient Engagement** — Predict and optimize engagement across digital health channels. — **+45%** Portal Engagement ### Capabilities - On-premise deployment to support HIPAA-aligned workflows - Clinical + behavioral data fusion - Temporal pattern analysis for disease progression - Patient similarity and cohort discovery - Wearable data integration - Event-level attribution for clinician review ### Results - **+29%** — Adherence — Improvement in medication adherence - **-31%** — Readmissions — Reduction in 30-day readmissions - **20+** — Healthcare Clients — Hospital systems and payers - **HIPAA-aligned** — Deployment — On-premise sovereign deployment supports HIPAA workflows ### Questions you can answer Zero-shot — no custom model training required. - "Which patients will miss their medications this week?" - "Who is at risk of readmission within 30 days?" - "What care pathway will produce the best outcome?" - "How should we personalize patient outreach?" - "Which patient segments need proactive intervention?" ### Case Study **Regional Hospital Network** - **Challenge** — High readmission rates and low medication adherence with no predictive capabilities. - **Solution** — Behavioral model trained on EHR, portal, and pharmacy interaction data across 2M patients. - **Result** — +29% adherence, -31% readmissions, on-premise HIPAA-aligned deployment in 6 weeks. ### Ready to transform healthcare? See how BaseModel.ai can answer your most critical healthcare questions — in days, not months. Book a Demo Explore Other Industries --- ## Insurance — https://basemodel.ai/industries/insurance All Industries Insurance ### Underwrite smarter and retain more policyholders Behavioral intelligence for claims prediction, fraud detection, pricing optimization, and policyholder retention. Book a Demo **+19%** — Loss Ratio Improvement ### Overview Insurance is fundamentally a behavioral prediction problem. BaseModel.ai enriches traditional actuarial models with behavioral signals to improve underwriting, detect fraud, and prevent lapses. ### Use Cases - **Claims Prediction** — Predict claim likelihood and severity using policyholder behavioral patterns. — **+19%** Loss Ratio - **Fraud Detection** — Identify fraudulent claims using behavioral sequence analysis and network patterns. — **94%** Detection Rate - **Lapse Prevention** — Predict which policyholders will lapse and trigger retention interventions. — **-27%** Lapse Rate - **Pricing Optimization** — Behavioral-enriched pricing models that better segment risk. — **+11%** Premium Accuracy ### Capabilities - Actuarial + behavioral data integration - Claims sequence pattern analysis - Agent and broker performance modeling - Regulatory-compliant explanations - Cross-product behavioral analysis - Real-time underwriting enrichment ### Results - **+19%** — Loss Ratio — Improvement through better risk prediction - **-27%** — Lapse Rate — Reduction in policy lapses - **15+** — Insurance Clients — Life, P&C, and health insurers - **94%** — Fraud Detection — With minimal false positives ### Questions you can answer Zero-shot — no custom model training required. - "Which policyholders are most likely to file a claim?" - "Is this claim potentially fraudulent?" - "Which policies are at risk of lapsing?" - "What premium best reflects this applicant's risk?" - "Which agents are underperforming?" ### Case Study **Multi-Line Insurer** - **Challenge** — Separate claims, fraud, and retention models with limited behavioral data utilization. - **Solution** — Single model incorporating all policyholder interactions across product lines. - **Result** — +19% loss ratio improvement, -27% lapse rate, deployed across 3 product lines. ### Ready to transform insurance? See how BaseModel.ai can answer your most critical insurance questions — in days, not months. Book a Demo Explore Other Industries --- ## Gaming — https://basemodel.ai/industries/gaming All Industries Gaming ### Keep players engaged and spending Predict player behavior — from session length to purchase intent to churn risk — using a single behavioral foundation model. Book a Demo **+38%** — Player Retention ### Overview Gaming generates some of the richest behavioral data of any industry. BaseModel.ai captures every in-game action, social interaction, and purchase to predict and optimize player engagement. ### Use Cases - **Player Retention** — Identify players about to churn and trigger personalized re-engagement. — **+38%** D30 Retention - **Monetization** — Predict purchase intent and optimize offer timing, pricing, and content. — **+45%** ARPDAU Lift - **Matchmaking** — Behavioral-based matchmaking that maximizes engagement and satisfaction. — **+22%** Session Length - **Content Recommendation** — Recommend game content, modes, and features based on play style embeddings. — **+30%** Feature Adoption ### Capabilities - Real-time play session analysis - Social graph embedding - Progression pattern modeling - Whale prediction from early signals - Cross-game behavioral transfer - A/B test acceleration ### Results - **+38%** — D30 Retention — Day-30 retention improvement - **+45%** — ARPDAU — Revenue per daily active user - **20+** — Gaming Studios — From mobile to AAA - **1B+** — Events/Day — In-game events processed ### Questions you can answer Zero-shot — no custom model training required. - "Which players will stop playing this week?" - "When is the optimal time to show this player an offer?" - "What content will keep this player engaged?" - "Which new players will become high-value?" - "How should we match these players for optimal fun?" ### Case Study **Top Mobile Gaming Studio** - **Challenge** — Declining D30 retention and ineffective IAP offers based on simple rule-based segmentation. - **Solution** — Behavioral model trained on 1B+ daily in-game events across 50M players. - **Result** — +38% D30 retention, +45% ARPDAU, real-time personalization deployed in 2 weeks. ### Ready to transform gaming? See how BaseModel.ai can answer your most critical gaming questions — in days, not months. Book a Demo Explore Other Industries --- ## Automotive — https://basemodel.ai/industries/automotive All Industries Automotive ### Predict purchase timing and maximize dealership performance From online research to showroom visit to service appointment — understand every customer's automotive journey. Book a Demo **+33%** — Lead Conversion ### Overview Automotive has long, complex purchase cycles. BaseModel.ai captures every digital interaction, dealer visit, and service event to predict purchase timing and optimize the customer journey. ### Use Cases - **Purchase Timing** — Predict when each prospect will be ready to buy and which vehicle they'll choose. — **+33%** Lead Conversion - **Service Retention** — Predict service appointment compliance and trigger proactive outreach. — **+24%** Service Retention - **Trade-In Prediction** — Identify owners ready to trade in and pre-approve optimal offers. — **+19%** Trade-In Volume - **Dealer Performance** — Behavioral analytics for dealer network optimization. — **+15%** Close Rate ### Capabilities - Long-cycle purchase journey modeling - Online-to-offline attribution - Dealer network behavioral analysis - Service and warranty pattern prediction - Connected vehicle data integration - Multi-household decision modeling ### Results - **+33%** — Lead Conversion — Improvement in qualified lead conversion - **+24%** — Service Retention — Service department repeat visits - **15+** — OEM Clients — Global automotive brands - **50M+** — Customer Profiles — Active automotive customer profiles ### Questions you can answer Zero-shot — no custom model training required. - "When will this prospect be ready to purchase?" - "Which vehicle model best fits this customer?" - "Will this owner return for their next service?" - "Which trade-in offer will this owner accept?" - "Which dealers need performance intervention?" ### Case Study **Global Automotive OEM** - **Challenge** — Low lead conversion rates and poor service retention with fragmented customer data across dealer network. - **Solution** — Unified behavioral model across digital, dealership, and service touchpoints for 50M customers. - **Result** — +33% lead conversion, +24% service retention, deployed across 2,000 dealers. ### Ready to transform automotive? See how BaseModel.ai can answer your most critical automotive questions — in days, not months. Book a Demo Explore Other Industries --- ## Software / SaaS — https://basemodel.ai/industries/software All Industries Software / SaaS ### Convert more trials and expand accounts Predict trial-to-paid conversion, expansion opportunities, and churn risk from product usage behavioral patterns. Book a Demo **+41%** — Trial Conversion ### Overview SaaS companies have rich product usage data but struggle to turn it into predictions. BaseModel.ai transforms every click, feature usage, and support interaction into actionable signals. ### Use Cases - **Trial Conversion** — Predict which trial users will convert based on early product engagement patterns. — **+41%** Conversion Rate - **Expansion Revenue** — Identify accounts ready for upsell or cross-sell based on usage trajectory. — **+29%** Net Revenue Retention - **Churn Prevention** — Detect declining engagement patterns and trigger customer success interventions. — **-36%** Logo Churn - **Product-Led Growth** — Optimize onboarding flows and feature discovery based on behavioral segments. — **+25%** Activation Rate ### Capabilities - Product usage behavioral modeling - Feature adoption sequence analysis - Account health scoring - Onboarding funnel optimization - Support interaction correlation - Multi-product usage fusion ### Results - **+41%** — Trial Conversion — Improvement in trial-to-paid conversion - **+29%** — NRR — Net revenue retention improvement - **35+** — SaaS Clients — From startup to enterprise - **10M+** — Users Modeled — Active SaaS user profiles ### Questions you can answer Zero-shot — no custom model training required. - "Will this trial user convert to paid?" - "Which accounts are ready for expansion?" - "Is this account at risk of churning?" - "What onboarding step is causing drop-off?" - "Which features drive the most retention?" ### Case Study **Series C SaaS Platform** - **Challenge** — 12% trial conversion rate with no predictive model for identifying high-intent users. - **Solution** — Behavioral model trained on product usage events across 500K monthly active users. - **Result** — +41% trial conversion, +29% NRR, deployed in 8 days. ### Ready to transform software / saas? See how BaseModel.ai can answer your most critical software / saas questions — in days, not months. Book a Demo Explore Other Industries --- ## Payments — https://basemodel.ai/industries/payments All Industries Payments ### Authorize smarter and prevent fraud at scale Real-time behavioral scoring for transaction authorization, fraud prevention, and merchant risk management. Book a Demo **99.9%** — Authorization Accuracy ### Overview Payment processors handle millions of transactions per second. BaseModel.ai delivers sub-millisecond behavioral scoring that improves authorization rates while reducing fraud losses. ### Use Cases - **Transaction Scoring** — Real-time behavioral risk scoring for every transaction at sub-millisecond latency. — **99.9%** Auth Accuracy - **Fraud Prevention** — Behavioral anomaly detection that catches sophisticated fraud patterns. — **-72%** Fraud Losses - **Merchant Risk** — Predict merchant risk and compliance issues from transaction behavioral patterns. — **96%** Risk Detection - **Revenue Optimization** — Reduce false declines to maximize authorization rates and revenue. — **+8%** Auth Rate ### Capabilities - Sub-millisecond transaction scoring - Network-level behavioral analysis - Merchant behavioral profiling - Cross-border transaction patterns - Velocity and pattern detection - PCI-compliant deployment ### Results - **99.9%** — Auth Accuracy — Transaction authorization accuracy - **-72%** — Fraud Losses — Reduction in fraud-related losses - **10+** — Payment Clients — Processors and networks - **5B+** — Daily Transactions — Scored in real time ### Questions you can answer Zero-shot — no custom model training required. - "Should this transaction be authorized?" - "Is this transaction pattern fraudulent?" - "What is this merchant's risk profile?" - "Which false declines can we recover?" - "What emerging fraud patterns should we watch?" ### Case Study **Global Payment Processor** - **Challenge** — Rule-based fraud system with high false decline rate costing $200M annually in lost revenue. - **Solution** — Behavioral foundation model scoring 5B daily transactions with contextual risk assessment. - **Result** — 99.9% auth accuracy, -72% fraud losses, +8% authorization rate. ### Ready to transform payments? See how BaseModel.ai can answer your most critical payments questions — in days, not months. Book a Demo Explore Other Industries --- ## Cybersecurity — https://basemodel.ai/industries/security All Industries Cybersecurity ### Detect threats through behavioral intelligence Transform security logs and user behavior into real-time threat detection — catching attacks that signature-based systems miss. Book a Demo **97%** — Threat Detection Rate ### Overview Security teams are overwhelmed by alerts. BaseModel.ai learns normal behavioral patterns and detects anomalies that indicate real threats — insider attacks, compromised accounts, and advanced persistent threats. ### Use Cases - **User Behavior Analytics** — Detect compromised accounts and insider threats through behavioral anomaly detection. — **97%** Detection Rate - **Threat Intelligence** — Correlate behavioral signals across the kill chain to identify advanced threats. — **-80%** Alert Fatigue - **Access Risk Scoring** — Real-time risk scoring for every authentication and access request. — **Sub-ms** Scoring Latency - **Incident Prioritization** — Automatically prioritize security incidents based on behavioral context and impact. — **+65%** SOC Efficiency ### Capabilities - User and entity behavioral analytics (UEBA) - Network traffic behavioral modeling - Authentication pattern analysis - Kill chain correlation - Insider threat detection - Zero-trust enrichment ### Results - **97%** — Threat Detection — Detection rate for advanced threats - **-80%** — Alert Fatigue — Reduction in false positive alerts - **15+** — Security Clients — Enterprise and government - **Sub-ms** — Scoring — Real-time risk assessment ### Questions you can answer Zero-shot — no custom model training required. - "Is this user account behaving abnormally?" - "Which access requests represent the highest risk?" - "What is this attack's behavioral signature?" - "Which alerts are real threats vs. noise?" - "Is there an insider threat in this department?" ### Case Study **Fortune 500 Enterprise** - **Challenge** — SOC team drowning in 10,000+ daily alerts with 95% false positive rate. - **Solution** — Behavioral baseline model for 100K employees with real-time anomaly scoring. - **Result** — 97% threat detection, -80% false positives, +65% SOC efficiency. ### Ready to transform cybersecurity? See how BaseModel.ai can answer your most critical cybersecurity questions — in days, not months. Book a Demo Explore Other Industries --- ## Fashion — https://basemodel.ai/industries/fashion All Industries Fashion ### Predict trends and personalize style Understand every customer's style preferences, purchase patterns, and trend sensitivity to drive personalized fashion experiences. Book a Demo **+35%** — Full-Price Sell-Through ### Overview Fashion retail combines complex trend dynamics with individual style preferences. BaseModel.ai captures both to optimize merchandising, personalization, and inventory decisions. ### Use Cases - **Style Personalization** — Recommend products that match each customer's evolving style profile. — **+35%** Sell-Through - **Trend Prediction** — Detect emerging style trends from behavioral signals before they peak. — **3 weeks** Early Detection - **Size & Fit** — Predict optimal sizing recommendations to reduce returns. — **-25%** Return Rate - **Markdown Optimization** — Optimize markdown timing and depth for each product-segment combination. — **+12%** Margin Lift ### Capabilities - Style preference embedding - Trend velocity detection - Size affinity modeling - Visual + behavioral similarity - Seasonal pattern learning - Cross-category discovery ### Results - **+35%** — Sell-Through — Full-price sell-through improvement - **-25%** — Returns — Reduction through better recommendations - **20+** — Fashion Clients — Luxury to fast fashion - **3 weeks** — Trend Lead — Average early trend detection ### Questions you can answer Zero-shot — no custom model training required. - "What style does this customer prefer?" - "Which trends are emerging in this market?" - "What size should we recommend?" - "When should we markdown this product?" - "Which customers will respond to this collection?" ### Case Study **Premium Fashion Brand** - **Challenge** — High return rates (30%) and poor full-price sell-through due to generic recommendations. - **Solution** — Behavioral style embeddings across browsing, try-on, purchase, and return data. - **Result** — +35% sell-through, -25% returns, deployed across 200 stores and online. ### Ready to transform fashion? See how BaseModel.ai can answer your most critical fashion questions — in days, not months. Book a Demo Explore Other Industries --- ## Customer Service — https://basemodel.ai/industries/customer-service All Industries Customer Service ### Resolve issues before customers call Predict escalations, optimize routing, and personalize support — transforming reactive service into proactive engagement. Book a Demo **-45%** — Escalation Rate ### Overview Customer service generates rich behavioral signals. BaseModel.ai transforms support interactions, product usage, and satisfaction patterns into predictions that prevent issues and improve resolution. ### Use Cases - **Escalation Prevention** — Predict which tickets will escalate and intervene with optimal resolution strategies. — **-45%** Escalations - **Smart Routing** — Route tickets to the best agent based on issue behavioral patterns and agent expertise. — **+28%** FCR Rate - **Proactive Outreach** — Detect product issues from behavioral signals before customers contact support. — **-35%** Inbound Volume - **Agent Performance** — Model optimal resolution behaviors and coach agents with behavioral insights. — **+22%** CSAT Lift ### Capabilities - Ticket sequence pattern analysis - Agent-customer matching optimization - Product issue early detection - Sentiment trajectory modeling - Multi-channel interaction fusion - Resolution path optimization ### Results - **-45%** — Escalations — Reduction in ticket escalations - **+28%** — FCR — First contact resolution improvement - **25+** — Service Clients — Enterprise contact centers - **-35%** — Inbound Volume — Through proactive issue detection ### Questions you can answer Zero-shot — no custom model training required. - "Will this ticket escalate?" - "Which agent is best for this issue?" - "Which customers are about to contact support?" - "What resolution will satisfy this customer?" - "Which product issues are emerging?" ### Case Study **Enterprise SaaS Provider** - **Challenge** — High escalation rate and low first-contact resolution across 500-agent contact center. - **Solution** — Behavioral model trained on 5M support interactions and product usage data. - **Result** — -45% escalations, +28% FCR, +22% CSAT, deployed in 3 weeks. ### Ready to transform customer service? See how BaseModel.ai can answer your most critical customer service questions — in days, not months. Book a Demo Explore Other Industries --- ## News & Publishing — https://basemodel.ai/industries/news-publishing All Industries News & Publishing ### Maximize reader engagement and subscription conversion Predict content preferences, optimize paywalls, and personalize the reading experience at scale. Book a Demo **+52%** — Subscription Conversion ### Overview Publishers sit on rich behavioral data — reading patterns, scroll depth, sharing behavior, subscription interactions. BaseModel.ai transforms this into predictions that drive engagement and revenue. ### Use Cases - **Paywall Optimization** — Predict the optimal moment and content to trigger a paywall for each reader. — **+52%** Conversion - **Content Recommendation** — Personalize content feeds based on reading behavior, not just topic preferences. — **+38%** Engagement - **Subscriber Retention** — Identify at-risk subscribers and trigger re-engagement campaigns. — **-30%** Churn Rate - **Ad Optimization** — Maximize ad revenue through behavioral targeting and viewability prediction. — **+25%** Ad Revenue ### Capabilities - Reading behavior sequence modeling - Scroll depth and engagement scoring - Content affinity embedding - Paywall timing optimization - Newsletter personalization - Cross-platform reader identity ### Results - **+52%** — Sub Conversion — Paywall conversion improvement - **+38%** — Engagement — Content engagement increase - **15+** — Publishers — National and regional outlets - **200M+** — Readers — Monthly active readers modeled ### Questions you can answer Zero-shot — no custom model training required. - "When should we show this reader a paywall?" - "What content will this reader engage with next?" - "Is this subscriber about to cancel?" - "Which newsletter topics maximize opens?" - "What ad placements maximize viewability?" ### Case Study **National News Publisher** - **Challenge** — Flat subscription growth with generic paywall rules showing to all readers at the same threshold. - **Solution** — Behavioral model personalizing paywall timing and content for 50M monthly readers. - **Result** — +52% subscription conversion, +38% engagement, -30% subscriber churn. ### Ready to transform news & publishing? See how BaseModel.ai can answer your most critical news & publishing questions — in days, not months. Book a Demo Explore Other Industries --- # Integrations — https://basemodel.ai/integrations Integrations ## Connect your warehouse, deploy your way BaseModel.ai meets your data where it lives. Install as a Snowflake Native App in one click, or deploy the GPU container on your own infrastructure — with seven first-class warehouse and lakehouse connectors either way. ## Supported Integrations ### Snowflake Install BaseModel.ai as a Snowflake Native App from the Marketplace. Train private behavioural foundation models in a built-in Web UI inside Snowsight — your data never leaves your Snowflake account. - Native App on the Snowflake Marketplace - 30-day free trial · billed through Snowflake - In-Snowflake Web UI · no docker, no YAML - Snowflake security review completed ### Docker Deploy BaseModel.ai as a GPU-enabled container on your own infrastructure. Run interactively for exploration or in CLI mode for scheduled training and scoring jobs. - Interactive & CLI modes - NVIDIA GPU runtime - Shared-memory tuning - Reproducible YAML pipelines ### Data Connectors Seven first-class data connectors cover the warehouses, lakehouses, and table formats most enterprise behavioral data sits in. Each connector is configured through the same YAML data block. - Snowflake & BigQuery - Azure Synapse & Databricks - Hive & ClickHouse - Local & cloud Parquet ### Prediction Outputs Predictions are written back to your warehouse as tables, or emitted as TSV files for downstream systems — CDPs, marketing automation, BI dashboards, or activation pipelines you already operate. - Warehouse write-back - TSV file export - One row per entity - Score + metadata columns ## Configuration via YAML Every BaseModel.ai pipeline — data, model, training, scoring — is a single YAML file. Version-controllable, environment-overridable, reproducible across runs. Below is the data block for a Snowflake source; swap `data_location.database_type` from `snowflake` to `bigquery`, `databricks`, `parquet`, or any of the seven supported connectors and the rest of the pipeline is unchanged. - Declarative — the YAML is the contract - Git-friendly, diff-friendly - Multi-table, multi-entity supported - Per-environment overrides ## Need a custom integration? If your data sits behind a connector we don't ship, our team can scope a custom adapter. Reach out and we'll figure out the best path. Contact Us --- # Snowflake Native App — https://basemodel.ai/integrations/snowflake Back to Integrations Snowflake Native App ## BaseModel.ai for Snowflake Install from the Snowflake Marketplace and start training private behavioural foundation models the same day. Everything runs inside your Snowflake account — your data never leaves, and Synerise never sees it. - Snowflake security review completed - No data egress - 30-day free trial - Billed through Snowflake Get on Snowflake Marketplace Read the Quick Start Watch the demo ## From Marketplace to predictions in four steps The Native App takes care of the infrastructure. You point it at your data and use the built-in UI. ### Step 1 — Install from the Snowflake Marketplace Open Snowsight, go to Data Products → Apps, search BaseModel.ai by Synerise, and click Get. No procurement, no contracts, no IT ticket required to start — a 30-day free trial unlocks immediately. ### Step 2 — Snowflake provisions everything for you The Native App spins up its own GPU compute pool, storage, image repository, and networking inside your Snowflake account — automatically, behind the scenes. There is no docker push, no YAML, no compute pool to size by hand, no service role to wire up. ### Step 3 — Train Foundation Models in the in-Snowflake Web UI BaseModel.ai ships its own UI inside Snowsight. Point it at your behavioural tables, configure the foundation model, and train. Each model has a clear status — Draft, Not Trained, Operational — with last-run / next-run timestamps and one-click re-runs. ### Step 4 — Fine-tune scenarios; scores land back in Snowflake Add scenario heads on top of the foundation — binary / multiclass / multilabel classification, regression, or recommendation. Define the target in Python, pick an audience filter, set a one-time or continuous training schedule. Predictions write back to your Snowflake tables. ## A full UI — inside Snowsight No notebooks to babysit, no YAML to hand-edit. Configure foundation models and scenarios in a purpose-built web app that runs inside your Snowflake account. ### Foundation Models — Manage every foundation model from one screen List, filter and search all foundation models trained in your account. Statuses (Draft / Not Trained / Operational), last-run and next-run timestamps, and a row-level menu to edit, schedule, or back up the configuration. - Filter by status: All · Draft · Not Trained · Operational - Last Run / Next Run / Edited timestamps per model - One-click New Foundation Model button - Configuration backups built in ### Fine-tuned Models (Scenarios) — Build and operate scenario heads visually A dedicated Scenarios screen lists every fine-tuned model with its type, status, upstream foundation, and schedule. Configuration is broken into clear blocks — Basic Settings, Target Function, Training Audience, and Training Schedule. - Statuses: Scheduled · Training · Ready · Training Failed · Draft - Types: Classification (Binary / Multiclass / Multilabel) · Regression · Recommendation - Target Function defined in a built-in Python editor - One-time or continuous training; auto-retrain when the upstream foundation updates ## Your data stays in your Snowflake The Native App framework is built around a single rule: the app runs inside your account. Synerise never gets a copy of your data, your model weights, or your predictions. ### Snowflake security review completed BaseModel.ai is published under the Snowflake Native Apps Framework — Secure by Design — and has passed Snowflake's automated security review process. ### Your data never leaves your Snowflake account The app runs entirely inside your account boundary on Snowpark Container Services. Synerise has zero visibility into your data, your tables, or the trained models — at any point in the lifecycle. ### Role-based access control Standard Snowflake RBAC governs which roles can install the app, grant it access to specific tables, and consume the resulting predictions. Nothing implicit, nothing global. ### Review every event you share The Native App declares which events it requires. You review and audit each grant in Snowsight before approving — and revoke any time. Sensitive-data protection is enforced by Snowflake itself, not by the provider. Read more about Snowflake's Native Apps Framework Secure by Design overview. ## Free to try. Billed through Snowflake. Start the same day you find the listing. Convert to a paid contract when you're ready — on a single Snowflake invoice. ### 30-day free trial Install from the Marketplace and get full, unlimited access for 30 days. You only pay the underlying Snowflake infrastructure cost (compute, storage) during the trial. ### Custom contract after the trial Continued use is covered by a custom contract negotiated directly with Synerise — sized to your data volume, scenario count, and SLA needs. ### Billed through Snowflake All charges flow through your Snowflake Marketplace bill. No second vendor invoice, no separate procurement cycle, no new payment relationship. ## Benchmarked against the state of the art The same scores published on the Marketplace listing — and the full benchmark suite is available on this site. **+292.4%** — vs SASRec (Meta HSTU comparison) Lift on sequential recommendation, with substantially faster training and inference than HSTU. **+61.57%** — vs TIGER Lift on sequential recommendation, without TIGER's expensive beam-search decoding step at inference. See the full benchmark suite ## Try BaseModel.ai inside your Snowflake One click from the Marketplace. 30 days free. Your data never leaves your account. Get on Snowflake Marketplace Talk to Sales Need a self-managed deployment instead? Snowpark Container Services tutorial covers the advanced bring-your-own-container path. --- # Docker / GPU — https://basemodel.ai/integrations/docker Back to Integrations Docker Deployment ## Deploy with Docker Run BaseModel.ai on your own GPUs as a Docker container. Interactive mode for exploration and configuration; CLI mode for scheduled training and scoring jobs you wire into your own orchestrator. ### Interactive Mode Explore and experiment Open a shell inside the container to inspect data, validate YAML configs, and run training commands by hand. Best fit for initial setup and debugging. - Full shell access - Inspect data and configs - Step-by-step training - Debug and iterate ### CLI Mode Run a single command Run a single training or scoring command and exit. Use it from cron, Airflow, Argo, or whatever scheduler you already operate — the container exits with a standard return code you can act on. - Single-command execution - Schedulable from any orchestrator - Config file is the contract - Standard exit codes ## Hardware Requirements The canonical hardware profile from the BaseModel.ai docs. Training and inference scale linearly with the number of GPUs. ### GPU - NVIDIA A100 recommended - Minimum: multi-GPU A10s / L40s - CUDA 12+ required - Multi-GPU scales linearly ### RAM & Disk - 240 GB system RAM - 1 TB disk for data + checkpoints - SSD strongly recommended - Mount data volumes for persistence ### CPU & Runtime - 32 CPU cores - Docker-capable environment - --shm-size=16g or higher - NVIDIA Container Toolkit ### Production performance profile From the BaseModel.ai docs Reference deployment from the canonical hardware appendix — ~8 B events / ~18 M unique clients / ~1 M products on a single NVIDIA A100. **12 h** — Foundation training (1× NVIDIA A100) **10 h** — Scenario fine-tuning (16,000 brands · 1× A100) **2,718** — Inference clients/sec/GPU (Full ranking throughput) **~8 B** — Events trained (~18 M clients · ~1 M products) ## Parquet YAML Config A complete Parquet-on-disk config: an event source joined to an attribute source via `join_on`, an entity-based 90/10 split with a fixed validation cutoff, and DDP training across two GPUs at `bf16-mixed`. Drop the same config into the container with a volume mount and you're training. - Same YAML schema across all 7 connectors - DDP strategy + device list - Joined attribute tables via join_on - Entity, time, or hybrid splits ## Ready to deploy? Reach out for the container image and onboarding tarball, or jump straight into the YAML reference. Get Started YAML Reference Hardware Reqs --- # Research — https://basemodel.ai/research Research ## Open research, real impact BaseModel.ai runs on **Cleora-NX** and **TREMDE** — Synerise's proprietary, internally maintained extensions of the open-source Cleora and EMDE algorithms (both published by Synerise at ICONIP 2021) — together with a **customized FFN backbone** purpose-built for behavioural sequences, carrying proprietary inductive biases tuned to event data and trained against a distributional-matching objective. **BaseModel ships with 5-years of proprietary, in-house improvements over prior open-source publications.** The published Cleora and EMDE papers describe the algorithmic foundations; Cleora-NX and TREMDE add multi-modality support, temporal-interaction handling, and significant performance and functionality improvements that are not described in any paper. No transformers, no attention, no recurrence. The full benchmark suite is described in the BaseModel.ai paper (preprint, ACM submission in flight). - **40+** — Synerise Publications - **3** — Foundation Papers - **4** — Top Venues ## Foundation Papers The three published Synerise papers describing the algorithmic foundations of BaseModel.ai — the BaseModel preprint plus the peer-reviewed Cleora and EMDE papers (ICONIP 2021). Production BaseModel runs Cleora-NX and TREMDE — proprietary, internally maintained extensions not described in these papers. ### BaseModel: A Foundation Model for Behavioral Data Foundation Model · Preprint · ACM submission in flight (venue TBD) · 2025 *Anonymous (under double-blind review)* The BaseModel.ai paper. Defines the foundation-model formulation for behavioural event streams, the distributional-matching pre-training objective, and the benchmark suite (sparse Amazon, Amazon Books, MovieLens, RelBench, rel-avito) on which BaseModel is reported to match or exceed every baseline tested. Production BaseModel runs Cleora-NX, TREMDE, and a customized FFN backbone — proprietary internal extensions of the published Cleora and EMDE foundations, with temporal-interaction modelling, multi-modality support, and other improvements that are not described in the paper. ### Cleora: A Simple, Strong and Scalable Graph Embedding Scheme Graph Embeddings · ICONIP 2021 (Springer LNCS, pp. 338–352) · 2021 *Rychalska B., Bąbel P., Gołuchowski K., Michałowski A., Dąbrowski J., Biecek P.* The published Cleora paper. Defines the deterministic, parameter-free hypergraph-embedding update T_{k+1} = normalize(P · T_k) and shows it is competitive with — or stronger than — learning-based graph-embedding baselines while running in time proportional to the number of hyperedges. The open-source predecessor of Cleora-NX, the proprietary, internally maintained engine that powers BaseModel.ai in production. BaseModel does not ship the published version; Cleora-NX adds multi-modality support, temporal-interaction handling, and significant performance and functionality extensions that are not described in this paper. ### An Efficient Manifold Density Estimator for All Recommendation Systems Density Sketches · ICONIP 2021 (Springer LNCS, pp. 323–337) · 2021 *Dąbrowski J., Rychalska B., Daniluk M., Basaj D., Gołuchowski K., Bąbel P., Michałowski A., Jakubowski A.* The published EMDE paper. Introduces compact density sketches whose sparse codes compose under summation, and shows that linear models on top of these sketches are competitive with deep recommenders on standard benchmarks while keeping inference cheap. The open-source predecessor of TREMDE, the proprietary, temporally-aware density-sketch engine inside BaseModel.ai. BaseModel does not ship the published version; TREMDE adds temporal-interaction modelling, modality-specific extensions, and other internal improvements that are not described in this paper. ## Additional Reading Engineering deep-dives, independent reproductions, and open-source code from the Synerise applied-AI research team that go beyond the published papers above. ### RecSys Challenge 2025: Universal Behavioral Profiles for Recommender Systems Universal Profiles · Proceedings of the Nineteenth ACM Conference on Recommender Systems (RecSys '25), pp. 1389–1393 · 2025 *Dąbrowski J., Janicka M., Sienkiewicz Ł., Stomfai G., Jannach D., Barile F., et al.* Synerise's RecSys Challenge 2025 paper, organised around the BaseModel concept of a single Universal Behavioral Profile per user. Defines the challenge protocol, the multi-task evaluation suite, and the profile-as-input formulation that BaseModel.ai productionises. ### The SYNERISE Dataset: An E-Commerce Dataset for Sequential Recommendation, Universal Behavior Modeling and Deep Relational Learning Dataset · Proceedings of the Recommender Systems Challenge 2025, pp. 1–6 · 2025 *Dąbrowski J., Janicka M., Sienkiewicz Ł., Stomfai G., Dietmar J., Barile F., et al.* The companion dataset paper for RecSys Challenge 2025: a large-scale, real-world e-commerce event log released by Synerise to benchmark sequential recommendation, universal behavior modeling, and relational deep learning under a common protocol. ### Synerise Monad: A Foundation Model for Behavioral Event Data Foundation Model (early) · Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR '23) · 2023 *Rychalska B., Łukasik S., Dąbrowski J.* Monad — the internal codename for early BaseModel.ai. Defines the foundation-model formulation for behavioral event streams and demonstrates that a single pretrained backbone transfers across downstream tasks (recommendation, churn, propensity) without retraining. ### The Monad Platform — Temporal Aspects in Behavioral Modeling Temporal Modeling · ECAI 2023, pp. 3226–3232 · 2023 *Rychalska B., Sieradzki I., Dąbrowski J.* How Monad (early BaseModel) handles temporal event streams: time-aware density-sketch encoding, time-conditioned modelling of behaviour, and the resulting gains on time-sensitive recommendation tasks. Direct precursor of the temporal-interaction handling that ships in TREMDE today. ### Synerise Monad — Real-Time Multimodal Behavioral Modeling Real-Time Multimodal · Proceedings of the 31st ACM International Conference on Information and Knowledge Management (CIKM '22) · 2022 *Dąbrowski J., Rychalska B.* Monad as a real-time, multimodal behavioral model: sketch-based profiles supporting sub-millisecond scoring across text, image, and interaction modalities. Establishes the multi-modal foundation that Cleora-NX and TREMDE later extend. ### T-EMDE: Sketching-Based Global Similarity for Cross-Modal Retrieval Density Sketches · arXiv preprint arXiv:2105.04242 · 2021 *Rychalska B., Wieczorek M., Dąbrowski J.* T-EMDE — the published precursor of TREMDE. Extends EMDE's density sketches with a temporally-aware similarity head and shows that the resulting sparse codes still compose under summation while improving cross-modal retrieval. ### Temporal Graph Models Fail to Capture Global Temporal Dynamics Methodology · arXiv preprint arXiv:2309.15730 · 2023 *Daniluk M., Dąbrowski J.* Empirical study showing that mainstream temporal graph neural networks miss global temporal dynamics that simple sketch-based models capture. Part of the evidence base behind BaseModel's choice to avoid transformers, attention, and recurrence in favor of Cleora-NX + TREMDE + a customized FFN backbone. ## Venues We Publish At - NeurIPS - KDD - ICONIP - ACM RecSys ## Join Our Research Team We're hiring researchers and engineers working on foundation models for behavioral data, graph embeddings, and density estimation. View Openings --- # About — https://basemodel.ai/about About ## Building the future of behavioral AI We're a team of researchers, engineers, and industry experts building one of the first behavioral foundation models. Based in Krakow, operating globally. ## Our Mission Our mission is to understand each and every action as accurately as possible from all available data. To predict behaviors. To make the right choices. To help you care. Join innovative organizations in using BaseModel.ai (Synerise) as a core AI framework to query the future with multiple questions in such segments like: retail, telco, finance, and more. - **150+** — Team Members - **32** — Countries - **18+** — Industries - **40+** — Research Papers ## Our Values ### Research-Driven Every product decision is rooted in published Synerise research (Cleora and EMDE at ICONIP 2021) and the proprietary internal extensions built on top of it (Cleora-NX, TREMDE, and the customized FFN backbone) — validated empirically on customer data. ### Customer-Centric We measure success by the measurable impact we create for our customers' businesses. ### Open Science We publish our research, open-source our tools, and contribute to the AI community. ### Diverse Thinking Our team spans 15+ nationalities, bringing diverse perspectives to complex problems. ## Our Journey - **2014** — Synerise Founded: Started as an AI-first customer data platform in Krakow, Poland. - **2018** — Cleora.ai Released: Open-source graph embedding engine, adopted by researchers worldwide. - **2020** — Behavioral AI Research: Published seminal papers on self-supervised learning from behavioral data. - **2022** — Foundation Model Concept: First prototype of behavioral foundation model trained on 10B+ events. - **2023** — BaseModel.ai Launched: Production-ready behavioral foundation model serving enterprise customers. - **2024** — Reference Production Deployment: Single-customer reference run from the BaseModel.ai docs: ~8 B events, ~18 M unique clients, ~1 M products — foundation training in 12 h on 1× NVIDIA A100. - **2025** — Global Expansion: Operating across 32 countries with sovereign cloud options. ## Headquartered in Krakow With offices across Europe and remote team members worldwide. We're always looking for exceptional talent. View Careers Contact Us --- # FAQ — https://basemodel.ai/faq ## Frequently Asked Questions Everything you need to know about BaseModel.ai — from architecture and deployment to governance and pricing. ## General ### What is BaseModel.ai? BaseModel.ai is a behavioral foundation model — a single model that ingests raw behavioral event data (clicks, purchases, sessions, transactions) from your data warehouse and learns the behavioral patterns of your population. Once it's pre-trained, lightweight scenario heads turn that shared representation into predictions for any individual-level question you can write a target function for — churn, fraud, purchase propensity, employee attrition, patient readmission, and more — in hours rather than the weeks a from-scratch model would take. The number of predictions scales without scaling your team. ### How is BaseModel.ai different from ChatGPT, Claude, or other LLMs? They solve fundamentally different problems. LLMs process and generate text — summaries, code, analysis. BaseModel.ai processes behavioral events and produces scored predictions on entities. The structures are different too: an LLM operates within a token context window; BaseModel ingests warehouse-scale tables spanning years of history across millions of entities. An LLM can help a data scientist write code for a single model; BaseModel learns the behaviors of your entire population once and then answers many individual-level questions from that same representation. The best an LLM can do is replicate a traditional data-science workflow — inspect data, create features, train an XGBoost model, tune hyperparameters. That's useful, but it produces a single-purpose model that doesn't benefit from the cross-domain behavioral knowledge BaseModel learns. ### How is a foundation model different from traditional ML? Traditional ML requires building separate models for each business question — each needing weeks to months of feature engineering, training, and maintenance. A foundation model is pre-trained once on all your behavioral data, learning a shared representation. New questions then attach as lightweight scenario heads on top of that representation, eliminating the need for separate end-to-end pipelines per task. ### What data does BaseModel.ai use? BaseModel.ai ingests raw behavioral event data — any timestamped interaction between entities. Website clicks, purchases, app events, transactions, support tickets, login events, and so on. The model also accepts attribute tables (product metadata, user profiles, content features) joined to the event stream. You don't need to pre-engineer features — the model learns from the sequence and context of events directly. ### Do I need a data science team to use BaseModel.ai? You don't need one to operate it day-to-day, but you do need a data engineer or ML practitioner to do the initial setup — wiring connectors, writing the YAML config, defining target functions, and connecting predictions back into downstream systems. Once that's in place, defining a new prediction is a small Python target function, not a new model build. ## Technology ### What is the architecture? Is it a transformer? No — and that's a deliberate design choice. BaseModel.ai is built from three proprietary components: Cleora-NX (Synerise's proprietary, internal extension of the open-source Cleora algorithm published at ICONIP 2021) produces hypergraph embeddings of your behavioural graph with multi-modality and temporal-interaction extensions; TREMDE (Synerise's proprietary, temporally-aware extension of the open-source EMDE / Efficient Manifold Density Estimator, also ICONIP 2021) compresses each entity's history into fixed-size density sketches whose shape the pipeline auto-tunes from your data, with temporal awareness; and a customized FFN backbone — purpose-built for behavioural sequences, with proprietary inductive biases tuned to event data — predicts the future sketch. No attention, no recurrence, but the backbone is engineered specifically for behavioural sequences and is not a textbook MLP. BaseModel ships with 5-years of proprietary, in-house improvements over prior open-source publications; Cleora-NX and TREMDE add significant performance and functionality improvements that are not described in the published papers. The training objective is distributional matching — cross-entropy between predicted and true sketch distributions across depths and modalities. ### What are behavioral sketches? Sketches are TREMDE's compressed representation of an entity's history. TREMDE — Synerise's proprietary, temporally-aware extension of the open-source EMDE algorithm (ICONIP 2021) — produces sparse codes that are composable under summation: adding two entities' codes yields the code for their combined behaviour, which is what makes per-user profiles cheap to build incrementally. TREMDE adds temporal-interaction modelling and modality-specific extensions not in the published version. Each modality — graph embeddings, text, images, tabular — is sketched independently and concatenated. Profiles are built on the fly with cost linear in interactions, and new items with content features get meaningful sparse codes immediately, so cold-start works without retraining. ### What is zero-shot vs scenario fine-tuning? After foundation pre-training, the model has a shared behavioral representation across your entire entity base. A scenario head is a small task-specific layer trained on top of that representation — it reuses all the heavy lifting the foundation already did. For most predictions, training a scenario head takes hours rather than the weeks a from-scratch model would take, because the representation is already there. ### What is Cleora — and how does it relate to what BaseModel actually ships? Cleora is the open-source graph embedding algorithm published by Synerise at ICONIP 2021. It treats each behavioural event as a hyperedge, builds a row-normalized transition matrix, and iteratively updates node embeddings via T_{k+1} = normalize(P · T_k) — deterministic (no training variance), parameter-free, and CPU-only. Importantly: BaseModel ships with 5-years of proprietary, in-house improvements over prior open-source publications. It runs Cleora-NX — Synerise's proprietary, internally maintained extension of Cleora that inherits the same deterministic update and adds multi-modality support, temporal-interaction handling, and significant performance and functionality extensions that are kept internal and not described in the published paper. The same relationship holds between the published EMDE algorithm and the proprietary TREMDE that runs inside BaseModel. ### What inference latency should I expect? Sub-millisecond for classification and regression scoring. For full-catalog recommendation scoring, the BaseModel research paper reports 9 ms per user to score the full 6 M-item rel-avito catalog on 1× H100. Production deployments tune for the latency budget the use case actually needs. ### What benchmarks has BaseModel been evaluated on? On RelBench (12 relational-database tasks across recommendation, classification, and regression) BaseModel matches or exceeds the best published baseline on 10 of 12 tasks. On sequential recommendation it improves over TIGER on Amazon Beauty / Sports / Toys by roughly +55–82 % across Recall@5, NDCG@5, Recall@10, and NDCG@10. On Amazon Books vs. HSTU-large it shows the largest gap — up to +138 % on NDCG@10. Full numbers (with means and standard deviations) are on the Benchmarks page. ## Deployment & Integration ### How long does deployment take? It depends on your data and your internal approvals more than on the model itself. The fastest path is: connect a warehouse, write the YAML data block, run foundation training, attach the first scenario head, write predictions back. With a clean warehouse and an A100 available, the first end-to-end prediction can land in under a week. Larger or more regulated environments take longer because of access reviews and data preparation, not because of the modeling step. ### What data sources can I connect? Seven first-class warehouse and lakehouse connectors: Snowflake, BigQuery, Azure Synapse, Databricks, Hive, ClickHouse, and Parquet (local or cloud). Each is configured through the same YAML data block — swap the database_type and connection_params and the rest of the pipeline is unchanged. Predictions can then be written back to the same warehouse as a table, or emitted as TSV files for downstream systems (CDPs, marketing automation, BI dashboards, activation pipelines) to consume. ### Can I deploy on-premise? Yes. BaseModel.ai ships as a Docker container with NVIDIA GPU runtime support, so you can run it on your own infrastructure or in a private VPC. Behavioral data never leaves the boundary you choose — when you run on Snowpark Container Services or your own Kubernetes, the model reads and writes within your account. ### How does BaseModel.ai integrate with my existing stack? Predictions land in your warehouse as tables (one row per entity, with a score and metadata columns) or as TSV files in a directory you choose. From there your existing CDP, marketing automation, BI, or activation pipelines pick them up the same way they consume any other warehouse table. There's no proprietary serving layer to integrate with. ## Governance & Security ### How is prediction quality monitored and audited? Standard metrics for each task type are produced at training and scoring time — AUC for binary, macro/micro F1 for multilabel, MAE/RMSE/R² for regression, MAP@K and NDCG@K for recommendation. Each scoring run is reproducible from the YAML config and pinned model version, so you can rerun any prediction batch and verify it matches what went out. Quality monitoring in production is set up the same way you'd monitor any other warehouse-resident model. ### How does BaseModel.ai handle bias and fairness? Fairness work is set up per deployment because the protected attributes, the relevant comparisons, and the regulatory constraints differ by use case. The pipeline supports excluding columns from training (via disallowed_columns), running standard fairness diagnostics on outputs, and slicing performance metrics by any dimension you ask the model to expose. We work with customers to scope what the right fairness audit looks like for their context. ### Is customer data protected? When BaseModel.ai runs on Snowpark Container Services or on your own GPUs, behavioral data never leaves your account boundary. The sketches the foundation model holds are aggregated, fixed-size representations rather than raw event logs, but they should still be treated as derivative of the underlying personal data and governed accordingly. PII columns can be excluded from training via the YAML config. ### What compliance certifications do you have? BaseModel.ai is delivered by Synerise, which operates under ISO/IEC 27001. The platform is designed to support GDPR, CCPA, HIPAA, and the EU AI Act controls customers need to put in place around behavioral models, but compliance is always a joint responsibility between us and the customer's deployment posture. For specific certification questions, please contact bm_support@synerise.com. ## Pricing & Support ### How is BaseModel.ai priced? BaseModel.ai is licensed per deployment. There is no public price sheet — please contact sales at bm_support@synerise.com to scope a license that matches your data volume, deployment topology, and support needs. ### Is there a free trial? We typically run a paid proof-of-concept where we deploy BaseModel.ai on a slice of your data and demonstrate results on agreed metrics. This lets you evaluate the platform on your own behavioral data before committing to a full deployment. Please contact us to scope one. ### What support is included? All commercial engagements include a named technical contact for onboarding, YAML config review, and production support. Larger deployments get a dedicated success manager and prioritized access to new connectors and modeling improvements. ### Still have questions? Our team is ready to help you scope a deployment on your own data. - Contact Us ---