Healthcare.AI
A clinical RAG system over FDA drug labels and ADA/JNC guidelines — hybrid retrieval, citation-grounded generation, and the evaluation pipeline that blocks low-faithfulness answers before deployment.
Role
Solo Project
Services
Hybrid RAG, LangGraph, Evaluation, Observability
Year
2025
Overview
ClinicalRAG answers clinical questions over FDA drug labels and ADA/JNC guidelines. In a medical context an ungrounded answer is not a bug, it is a hazard — so the system is built around proving its own reliability.
The gap between a RAG prototype and a production RAG system is measurement. Without faithfulness evaluation and drift detection you cannot distinguish a good answer from a lucky one, or tell whether a preprocessing change helped or hurt.
Approach
Engineered hybrid retrieval combining BM25 sparse indexing with dense embeddings and reranking, orchestrated through LangGraph with citation-grounded generation.
Designed custom preprocessing for healthcare PDFs: section-aware extraction, metadata preservation, and retrieval-aware chunking.
Ran retrieval ablation experiments across 9 chunking configurations — 3 strategies × 3 token sizes — to optimise context continuity for medical terminology.
Built an evaluation-driven pipeline with RAGAS, MLflow, and GitHub Actions CI to automate regression testing and block low-faithfulness responses before deployment.
Implemented custom drift detection and pipeline regression monitoring via Redis.
Outcome
Production-style latency SLAs defined and met: p50 <800ms, p95 <2.5s, p99 <5s.
Confidence scoring and observability hooks as reliability safeguards for clinical query workflows.