MediVise

Medical documents are difficult for patients to parse; the platform needed secure multi-tenant ingestion, hybrid retrieval over user-uploaded files, and latency-aware LLM responses without crossing into clinical advice.

Deliver a production-quality RAG backend for medical documents within a capstone scope—strict tenant isolation, controlled context windows, and bounded non-diagnostic AI outputs.

• Educational capstone scope—not a clinical or FDA-regulated product
• Per-user document isolation required at the database layer
• Latency-sensitive RAG over documents of varying quality and format

• Built a secure multi-tenant document ingestion API using Phi-4-mini, isolating user vectors and metadata with Supabase Row-Level Security over PostgreSQL.
• Improved retrieval quality with hybrid sparse/dense search, token-aware chunking, and caching for RxNorm lookups to reduce external API latency.

• Enforced tenant isolation with Supabase RLS policies rather than application-layer checks alone
• Used hybrid sparse/dense retrieval instead of semantic-only search to improve recall on medical terminology
• Aggressively cached RxNorm lookups to mitigate external API latency on the hot path

• Row-Level Security policies isolating user vectors and metadata per tenant
• Dynamic token-aware chunking to control retrieved context size
• Cached RxNorm lookups reducing redundant external API calls

Built a capstone RAG platform demonstrating secure multi-tenant document ingestion, hybrid retrieval tuning, and responsible AI boundaries in a healthcare-adjacent domain.

• Not shipped; educational capstone prototype only
• Not intended for diagnosis, treatment decisions, or medical advice

• Expand automated evaluation for RAG answer quality across document types
• Harden auth boundaries and document access policies before any broader release

• FastAPI
• PostgreSQL
• Supabase
• Phi-4-mini
• RAG
• Hybrid retrieval
• RxNorm API