gabrielpoku/Ghana-Medical-Agent

# Hybrid Fine-tuning and RAG for Medical QA in Low-Resource Settings (Ghana) A research project + deployable system: a phone-based medical triage assistant for Ghana that combines a **fine-tuned LLM** with **RAG** over Ghana's official clinical guidelines, with a **doctor-in-the-loop** web app for monitoring and human evaluation. > Triage + first-aid guidance only. NOT a diagnostic tool. See SAFETY.md. ## Repo layout ``` backend/ FastAPI service: Twilio voice webhook, safety layer, RAG, LLM inference webapp/ Next.js app: admin dashboard + doctor portal (Supabase) data/ Datasets — raw (STG PDF, etc.) and processed (chunks, splits) finetune/ Kaggle notebooks + scripts for QLoRA fine-tuning eval/ Benchmark + human-eval harness (AfriMed-QA) paper/ Manuscript, outline, figures scripts/ One-off utilities (PDF extraction, dataset download) ``` ## Planning docs - PLAN.md — full project plan, phases, architecture, paper outline. - DATA.md — datasets and their roles. - SAFETY.md — safety guardrails and escalation. - ETHICS.md — ethics, consent, privacy, IRB. ## Quick start (in progress) 1. `python scripts/extract_stg.py` — extract text from the Ghana STG 2017 PDF. 2. `python scripts/download_afrimedqa.py` — pull the AfriMed-QA benchmark. 3. See `finetune/README.md` for the Kaggle fine-tuning flow. 4. See `backend/README.md` to run the voice service locally. ## Status Phase 0–1 (scaffold + data). See PLAN.md for the phase roadmap.