Shubham Mishra

I built a hybrid dense + sparse RAG pipeline for searching internal documents, combining Qdrant dense indexing with FastEmbed (BM25) and Reciprocal Rank Fusion, then adding a cross-encoder reranker. On top of that, I shipped a FastAPI backend with JWT auth and SQLite3 query caching, packaged with Docker.

I also created an LLM prompt regression-testing system that runs 125 golden test cases per change and scores outputs using exact-match plus LLM-as-judge methods. It includes rolling-window drift detection and sends Slack alerts when output quality drops beyond a threshold.

My experience across the ML lifecycle includes preprocessing and evaluation workflows from a Deloitte virtual internship, plus exploring Azure Cognitive Services and AI deployment/lifecycle patterns during a Microsoft AINSI internship.