How it works
One spec. Three execution modes. Zero drift.
Define
Write transforms in Python or use the expression DSL for simple features. Group related features, set source dependencies, and register via API. Your features are versioned, searchable, and reusable.
Compute
Features run as actors on Apache Pekko with Python UDFs and an expression DSL. The same logic executes in real-time, on streams, or in batch. No code duplication, no training/serving skew.
Serve
Retrieve feature vectors through authenticated gRPC and HTTP paths. Feature values are tenant-scoped, observable, and backed by persisted provenance so operators can explain how a value was produced.
Three modes, one platform.
Whether you need features for a real-time API call, a streaming pipeline, or a historical backfill, the Enrichment Platform runs the same transformations across all three modes from a single definition.
Plan pilotReal-time. Authenticated feature reads for live product decisions
Stream. Updates from Kafka, Kinesis, CDC, and HTTP with persisted state
Batch. Historical feature calculation for training sets and backfills
Capabilities
From raw data to production features.
Feature specifications
- Declarative feature definitions with tenant ownership
- GitHub App onboarding for repository-managed bundles
- Plan/deploy evidence for each feature bundle change
- Versioned specs with dependency and provenance metadata
Transformation engine
- Python UDFs: write transforms in Python, executed sandboxed
- Expression DSL for config-driven features (no code deploy)
- Batch UDFs return multiple features from a single function
- Pre-installed numpy, pandas, scikit-learn for ML scoring
Feature store
- Elasticsearch-backed feature storage and retrieval
- Point-in-time feature lookups for training
- Entity-based and pipeline-based retrieval
- Automatic index rollover and retention
Serving & monitoring
- Authenticated gRPC and admin APIs
- Health checks and readiness probes
- OpenTelemetry tracing, Prometheus metrics, Grafana dashboards
- Horizontal autoscaling on Kubernetes
Tenant onboarding
Let customers manage features from their GitHub organization.
Tenant admins install a GitHub App, select repositories, link a branch and bundle path, then use plan/deploy actions in the platform. No customer GitHub Actions workflow is required for the core onboarding model.
Install
The platform creates a one-time state and sends the tenant admin to GitHub's installation screen for their organization.
Link
The tenant links a repository, branch, and bundle path containing datatier.json,
schemas, UDF files, and feature definitions.
Deploy
The platform fetches bundle contents through the GitHub App token, validates the plan, applies changes, and records deployment evidence.
Audit packages for regulated decisions.
Decision records are hash-chained per tenant and can be exported as sealed audit packages with manifests, row hashes, signatures, signing-key evidence, and offline verification.
See security modelRecord. Capture request context, feature values, source provenance, and actor metadata
Seal. Build tenant-local hash chains and sign export manifests
Verify. Auditors can validate package integrity without trusting the running platform
Enterprise readiness
Security, deployment, evidence, and pilot flow in one product view.
These are Enrichment Platform capabilities, not separate products. Review the control model, deployment shape, documentation evidence, and pilot checklist without leaving the product page.
Tenant isolation
Tenant API keys, role-scoped permissions, runtime tenant context, and PostgreSQL Row Level Security protect tenant-owned control-plane and audit data.
Audit integrity
Decision ledger rows are hash-chained per tenant. Export packages include manifests, row hashes, signatures, and signing-key evidence for independent verification.
UDF boundary
Python UDFs run through a worker boundary with explicit timeouts, dependency metadata, network posture, and operator controls for production sandbox mode.
Customer cloud
Deploy into a customer-controlled AWS account or a dedicated cloud environment. The platform is designed for private networking and customer-owned data stores.
Infrastructure layers
Terraform bootstraps cloud primitives. Helm installs the platform and foundation components such as ingress, secrets, observability, and policy controls.
Production validation
Staging should prove migrations, tenant onboarding, GitHub bundle sync, feature serving, audit export, failover drills, alerts, and load baseline before production rollout.
Onboarding runbooks
Tenant onboarding and GitHub App feature onboarding are documented as operator flows, API flows, and customer-facing docs-site entries.
Audit runbooks
Decision audit package generation, offline verification, access events, and scheduler evidence are part of the platform operations material.
API reference
Admin HTTP and gRPC APIs document tenant lifecycle, Git bundle deployment, feature operations, compliance exports, and operational endpoints.
Week 1: connect
Provision staging, create a tenant, configure auth, link GitHub, and register one source connector against non-production data.
Week 2: compute
Deploy a feature bundle, run feature queries, validate point-in-time behavior, and review operational metrics with the customer team.
Week 3: evidence
Export an audit package, verify signatures offline, review runbooks, and agree the production rollout criteria.
Built for SaaS platforms in regulated industries.
Most feature stores are designed for a single company's ML team. The Enrichment Platform is designed for SaaS companies that compute features per customer, with built-in tenant isolation, audit trails, and direct database connectivity.
Multi-tenant by design
Row-level security, per-tenant rate limiting, isolated feature computation, and cross-tenant sharing with access grants. Tenant-created resources stay scoped to the authenticated tenant.
Compliance-ready audit trail
Decision evidence is written to an append-only ledger with tenant-local hash chains. Export signed audit packages for regulatory review and incident investigation.
CDC: skip the middleware
Point a connector directly at your database or stream. Kafka is supported but not mandatory. Change Data Capture can turn database writes into feature updates.
Use cases
Feature infrastructure your compliance team will love.
Risk & underwriting
Compute risk features from claims history, telematics, and external data in real-time for instant decisions.
Fraud detection
Stream transaction data through feature pipelines. Flag anomalies with features computed from live and historical patterns.
Personalization
Build user profiles from behavioral streams. Serve real-time feature vectors to recommendation models.
Credit scoring
Combine bureau data, transaction history, and alternative data into consistent feature sets for model training and serving.
Pricing engines
Enrich pricing requests with computed features from multiple sources. Same features in batch training and real-time quoting.
ML training pipelines
Generate point-in-time correct training datasets with batch mode. Eliminate training/serving skew with unified feature definitions.
Architecture
Connects to your data stack.
- Kafka
- Kinesis
- S3
- CDC
- HTTP
- gRPC API
- Kafka
- S3 export
- Webhooks
Bring your own data sources. We handle the compute, storage, and serving.
Pilot path.
A useful pilot should validate tenant onboarding, one source connector, one Git-backed feature bundle, feature serving, and audit package export in a production-like staging environment.
Create tenant
POST /api/tenants
Link GitHub
POST /admin/git/github/sync
Export evidence
GET /admin/audit-packages
Pricing
Pay for what you compute.
Pilot and production pricing should align to deployment model, audit retention, support requirements, and feature computation volume.
Starter
Scoped
- One staging environment
- One tenant onboarding flow
- One source and feature bundle
- Audit export validation
Growth
Usage-based
- Production tenant onboarding
- GitHub App feature deployment
- Audit packages and retention controls
- Staging validation and runbooks
Scale
Custom
- Customer cloud or dedicated cloud deployment
- VPC peering / private endpoints
- SSO, KMS signing, object-lock exports
- Custom validation, SLA, and support model
Ready to validate tenant-scoped features?
Run a focused pilot that proves onboarding, feature deployment, serving, and audit evidence.