Adi Nayak

Work

User-Scoped Feature Flags

A config system for tagging individual users into new AI behavior, with cache strategy tuned to how critical each feature was.

Java / Kafka / Feature Flags / Caching

Brief

The team needed to test new AI behavior on live enterprise customers without risking a full deploy to everyone at once.

Approach

  • Targeted Rollouts: Built a config system and shared SDK for validating new AI behavior on individual users rather than a blanket percentage.
  • Tuned Caching: Allowed features that tolerate gradual consistency to read from a local cache for performance, while correctness-critical features bypassed it.
  • Instant Invalidation: Ensured toggling a feature invalidated the cache immediately, keeping behavior consistent across the fleet.

Result

Reduced the risk of shipping new AI behavior broadly by validating with individually tagged users first, with an instant kill switch if issues came up.