Bio

I build AI systems that keep working after the demo ends — not just wired up in a notebook, but running in production with real users and edge cases. What I bring: - AI agents & orchestration — LangChain, LangGraph, MCP tool integration, multi-agent workflows with conditional routing - RAG — hybrid dense + BM25 retrieval, reranking, vector databases (Milvus, pgvector) - Evaluation — RAGAS, LLM-as-a-judge, Langfuse/LangSmith tracing, so you know it's actually working, not just guessing - Fine-tuning — QLoRA/PEFT for domain-specific or structured-output tasks - Full-stack delivery — FastAPI + React + Docker, so you get a deployed system, not a proof of concept Selected projects: Vidhaan AI — Legal research platform Hybrid RAG search over 113 public acts (1,100+ sections), fusing dense + BM25 retrieval with reciprocal rank fusion. Added a multi-hop RAG layer over private per-user document vaults, so users query their own documents alongside public law in one search. 85ms retrieval latency at production scale. PriorAuth — Multi-agent healthcare workflow 5 specialized LangGraph agents (evidence extraction, compliance, risk assessment, decision synthesis) with human-in-the-loop checkpoints for high-risk cases. Full observability via Langfuse + LangSmith, tracing agent runs, latency, and cost. Delivered as a live dashboard, not a backend script. KAIROS — Fine-tuned prediction model Fine-tuned a Llama 3-8B model with QLoRA to output structured predictions from raw geopolitical news, replacing a zero-shot general-purpose model that couldn't reliably hit the required output format. Deployed as a live API on GCP cloud stack. Send me your use case and I'll tell you honestly whether an agent/RAG approach fits, or if something simpler will get you there faster.

Skills

Reviews (0)

No reviews yet.