Problem
A multimodal data pipeline broke under load and ran on tribal knowledge.
Kensho
Streaming data for financial research.
The solution
Led batch-to-streaming platform migration, hardware-aware routing, and durable on-call and onboarding systems.
Approach
Migrated batch ETL to Kafka and Faust streaming, running both pipelines in parallel until output matched before any cutover
Split GPU-intensive transcription from market-data ticks onto separate hardware-routed worker pools
Replaced tribal knowledge with explicit on-call ownership, escalation protocols, and self-describing alerts
Led the 15-person engineering org through the S&P Global acquisition integration
Next case
Humana
Care teams worked from fragmented, stale records. Connected care tools and brought claims data up to date.