Real-time ML Feature Store
Built a low-latency feature store powering risk models with millisecond SLAs and versioned features.
GoTypeScriptgRPCBigQueryRedis
Problem
Models starved for fresh features; batch pipelines led to stale predictions and drift.
Approach
Unified offline/online features, introduced stream processors, versioning, and consistency checks.
Architecture
Streaming ingestion (Kafka), Flink processors, Redis as online store, BigQuery as offline, gRPC APIs with typed contracts.
Tradeoffs
Higher infra cost and operational load for strong consistency and low latency guarantees.
Outcome
AUC up 7%, fraud losses down double digits; engineers self-serve features in hours not weeks.