At Brototype, I refactored an existing RAG chatbot codebase for retrieval quality and maintainability, and shipped approved changes to production. I also extended vector-only retrieval with BM25 and Reciprocal Rank Fusion.
I built DocMind, a multi-agent RAG document intelligence platform backed by PostgreSQL and pgvector. Its RAGAS evaluation across 50 queries recorded 0.93 faithfulness and 0.88 answer relevancy.
For DocMind, I designed a three-agent LangGraph supervisor system with LLM-as-a-Judge guardrails for prompt injection, PII leakage, and toxic content. I also built asynchronous ingestion and deployed the services on AWS.
My other projects include Delay2Decision, which uses machine learning for flight-delay prediction and anomaly detection, and FairCrop AI, whose LangGraph agents analyze agricultural markets and negotiate with buyers.
