Ankur Kumar

Headed the design and delivery of a multi-tenant featurization service that ingests large volumes of assets, such as images, video, and text, runs them through a configurable graph of feature extractors, and publishes a standardized feature contract for downstream ranking, search, and analytics.

I defined the ingestion and queuing model, built an orchestrator that explodes each asset into per-model tasks with dependencies, and implemented bounded-concurrency workers with back-pressure, timeouts, retries with exponential backoff, and circuit breakers to protect external model endpoints. I also introduced idempotency keys, an outbox/transactional log to achieve logical exactly-once behavior, and a dead-letter flow for irrecoverable cases.

I also designed the feature schema with versioning and lineage, such as model, version, parameters, and validation gates. I added row-level access controls and audit logs for tenant isolation.

Operationally, I set SLOs, dashboards, and alerts, tuned caching and batching to control costs, and ran load tests to size capacity.

The system handled 10,000+ assets per minute at peak with stable latency and eliminated duplicate model calls via deduplication.