Dev Sharma

I've shipped production GenAI systems at Astroraag, taking over a silently failing multi-agent LangGraph pipeline and rebuilding it into modular, observable agent nodes. The result made debugging faster for the team, reducing debug time by 60% and iteration time by 40%.

I optimize live customer-facing AI pipelines through prompt engineering, context engineering, token budgeting, response trimming, and multi-LLM experimentation. At Astroraag, I cut latency by 50% and operational cost by 70% while improving response consistency.

I built a hybrid RAG Knowledge QA system using QdrantDB, BM25, RRF, and CrossEncoder reranking, with query expansion and adaptive chunking for technical documentation. I also created retrieval evaluation workflows that helped improve retrieval coverage by about 40%.

Previously, I led a small team at SovereigntLabs to build an AI face-authentication system with DeepFace and Python. I enjoy turning vague product asks into deterministic architecture, from backend infrastructure and deployment to reliable LLM applications.