Built the behavioral matchmaking engine at the core of a production social platform: embedding-based profile scoring, a live match queue with presence-aware availability, and a timed decision window where both users choose to keep or release the match. Engineered for the race conditions that make live matchmaking hard: simultaneous accepts, mid-queue disconnects, and double-matching, resolved with atomic Redis operations. The engine learns from behavioral signals rather than stated preferences, with anti-bot and spam detection built into the pipeline.