At Lyra Health, I built Natural-Language Care Search, a two-stage recommendation system that improved search-to-booking conversion by 18%. I fine-tuned domain-specific embeddings to connect members’ everyday language with clinical terminology, lifting retrieval recall@50 by 31%.
I also developed the personalized ranking and constrained re-ranking stages, serving 150K+ ranked queries daily at p95 latency under 300ms. For Enhanced Triage, I shipped a model that reduced misrouted care episodes by 22% and cut time-to-first-appointment by 1.5 days.
At Amazon Web Services (AWS), I developed retrieval and ranking components for Amazon Bedrock Knowledge Bases and optimized model serving, reducing inference costs by about 35%. Earlier at edX, I designed a real-time adaptive recommender that improved recommendation CTR by 24% and course completion by 8%.
