Chalk

What Chalk does

Chalk is an AI/ML data platform that provides infrastructure for getting fresh, contextual data to models and agents at inference time, and for deploying and running those models and agents securely and at scale. It offers both feature/context and compute solutions for inference, allowing developers to define what they need in Python and run it on their own infrastructure. Chalk aims to replace a fragmented AI stack (feature store, vector database, retrieval/prompt tooling, orchestration, sandbox runtime, governance) with a single, unified platform. It ensures a single definition of data (features, embeddings, LLM outputs, prompts) as Python, computed once and served everywhere (training, real-time inference, agents). The platform guarantees point-in-time correctness, fast performance in production (single-digit millisecond latency), and deployment within the user's own cloud (VPC) for enhanced security and compliance. Chalk's architecture is designed to retrieve fresh features across heterogeneous sources with minimum latency, orchestrate complex transformations, and support online serving, offline training-set generation, and streaming ingestion in a single system, all while upholding enterprise-grade guarantees for security, observability, and reliability.

Products

• Context EngineA real-time feature computation engine that serves as the single source of truth for features. It allows defining features once in Python and computes and serves them directly from data sources at inference time in milliseconds. Unlike traditional feature stores, it keeps training and serving perfectly in sync, makes feature development faster, and includes built-in monitoring, branch-based deployment for feature versioning, and deployment within the user's cloud.
• Chalk ComputeExtends the platform to address context and infrastructure challenges of agents. It provides governed, sandboxed environments where agents securely run and take action on data without it leaving the user's VPC. This enables running scalable agents or remote processes quickly, managing data access, and evaluating agent performance against historical time periods. Sandboxes boot quickly and scale to many isolated containers, and can also run long-lived model servers, including GPU-backed inference servers.

Key capabilities

• Real-time serving engine
• High-volume workloads at ultra-low latency (100,000 QPS in 5ms)
• Perfect auditability and data replay
• Parallel Resolvers (Rust-based runtime)
• Unify training and serving
• Prevent train-serve skew
• Detect, troubleshoot, and eliminate data issues
• Observability (data use, drift, quality, lineage, audit logs, PII masking)
• Integrations with existing tools and infrastructure
• Deploy to your own infrastructure/cloud (AWS, GCP, Azure)
• Use existing database as online + offline store
• Single source of truth for features
• Compute features at query time (execution first)
• Point-in-time correctness
• Iterate fast with branches (branch deployments)
• Low-latency serving (sub-5ms)
• Discoverability & governance (built-in catalog, versioning, metadata)
• Generate point-in-time correct training datasets
• Temporal Aggregations (define once, reuse across batch, online, real-time)
• LLM Toolchain (unify prompt engineering, embeddings, vector search, inference)
• Declarative feature engineering in Python
• Sandboxed environments for agents (gVisor, scoped cloud identity, replayable)
• Automatic parallelism across pipeline stages
• Vectorization of pipeline stages
• Statistics-informed join planning
• Low-latency key/value stores (Redis, BigTable, DynamoDB)
• Transpilation of Python code into native code (symbolic Python interpreter)
• Dependency graph (DAG) of features
• Dynamic query plan generation
• Caching, incremental updates, backfills
• Scheduled queries
• Streaming ingestion (Kafka, Kinesis, Pub/Sub)
• Offline computation and serving (training data requests, point-in-time correct feature values)
• Integration with data providers (Snowflake, Delta Lake, BigQuery)
• Reverse ETL
• Feature Drift monitoring (Kolmogorov-Smirnov test, charts, alerts)
• Integration tests with `ChalkClient.check()`

Use cases

• Fraud & Risk (identity fraud detection, behavioral abuse, account risk, account takeover, onboarding & KYC, synthetic IDs & emerging fraud, transaction fraud, card authorization risk, account-level spend abuse, emerging payment fraud)
• Credit (income computation, credit scoring, credit underwriting, real-time credit approvals, buy now, pay later, cash advance, business lending, card issuing)
• Caching (cache busting)
• Predictive Maintenance (device data)
• Growth decisioning (personalize, predict, optimize revenue across customer lifecycles)
• Underwriting
• Search and ranking (adapt results instantly based on user context, behavior, intent)
• Recommender systems (serve recommendations based on real-time behavior and semantic similarity, live auction recommendations, ecommerce product/content feeds, financial services offers, marketplaces discovery, healthcare recommendations, embedding-based discovery)
• Risk decisioning
• Compliance
• Dynamic Pricing
• Routing and Optimization
• Resource Forecasting
• Anomaly Detection
• Transaction Scoring
• Contextual Decisioning
• Content Scoring
• Threat Monitoring
• ML development (experimentation, serving, governance)
• Personalization

AI approach

Chalk provides an AI/ML data platform that offers infrastructure for getting fresh, contextual data to models and agents at inference time, and for deploying and running those models and agents securely and at scale. It acts as a real-time feature computation engine and a unified platform for various ML components like feature stores, vector databases, and orchestration layers. Chalk allows users to define features and logic in Python, which are then computed and served in real-time. It supports both online serving and offline training-set generation, ensuring consistency between training and serving data. The platform also includes Chalk Compute for running agents and models in sandboxed environments within the user's cloud.

Tech named: Rust-based runtime, Python, Jupyter, XGBoost, SQL, gVisor, AWS, GCP, Azure, Redis, BigTable, DynamoDB, Kafka, Kinesis, Pub/Sub, Snowflake, Delta Lake, BigQuery, Kolmogorov-Smirnov test

Industries served

• Financial Services
• E-commerce
• Marketplaces
• Healthcare
• Energy

What it says sets it apart

• Context+ infra for agents + models
• Built for speed
• In your cloud (deploy to your own infrastructure/VPC)
• Uses existing database as online + offline store (no bespoke storage)
• Rust-based runtime for maximum performance
• Query data just-in-time for online predictions
• Perfect auditability and data replay
• Unifies training and serving to prevent train-serve skew
• Built-in observability for data issues, drift, and quality
• Single source of truth for every feature definition
• Computes features at query time rather than just storing values
• Point-in-time correctness for recomputing features as they would have been at inference time
• Fast iteration with branch deployments for experimentation
• LLM Toolchain for unifying prompt engineering, embeddings, vector search, and inference
• Governed, sandboxed environments for agents (Chalk Compute)
• Sandboxes boot in under a second and scale to 10,000 isolated containers in under 10 seconds
• Automatic parallelism, vectorization, and statistics-informed join planning
• Transpilation of Python code into native code with a symbolic Python interpreter
• Serves as a drop-in replacement for orchestration tools like Dagster, Airflow, and Prefect
• Directly interfaces with underlying data sources, managing connections and transformations
• Real-time feature computation for payments at authorization time
• Enterprise-grade security with deployment within your cloud, data residency, and IAM/networking integration
• Unified ML development, accelerating feature delivery and reducing time-to-production
• Centralized experimentation, serving, and governance for ML lifecycle
• Python-native feature development for seamless experimentation-to-production
• Feature store with built-in governance and versioning
• Ability to ingest multiple data sources (SQL, streams, Python UDFs, Expressions, APIs) while maintaining ultra-low response times

Funding rounds we track

From the AI funding tracker — rounds as reported by the linked publications.