Chezhira/daktari-ai

Offline-first, multilingual clinical AI for Community Health Workers in Africa — Gemma 4 + RAG over WHO & MSF guidelines # 🏥 Daktari AI ### Clinical Decision Support for Community Health Workers across Africa *Powered by Gemma 4 + RAG over WHO & MSF Clinical Guidelines* --- ## What is Daktari AI? "Daktari" means *Doctor* in Swahili. Daktari AI is an offline-first, multilingual AI assistant built for **Community Health Workers (CHWs)** — the frontline volunteers who are often the only point of healthcare access for millions of people across Sub-Saharan Africa. It uses **Gemma 4** (Google's open edge model) with **Retrieval-Augmented Generation (RAG)** over WHO and MSF clinical guidelines to provide: - **Grounded clinical guidance** — answers backed by real medical literature, not hallucinations - **Multilingual support** — English, Swahili, French, Hausa, Amharic, Portuguese - **Multimodal image analysis** — CHWs can photograph wounds, rashes, or malnourished children - **Offline-first operation** — runs on a cheap Android phone or laptop with no internet --- ## Architecture ``` [CHW types/speaks symptoms or uploads photo] ↓ Gemma 4 Vision (multimodal) ↓ Query → Embedding Model (all-MiniLM-L6-v2) ↓ ChromaDB Vector Store ← WHO IMCI Guidelines ← MSF Clinical Guidelines ← WHO Maternal & Newborn Care ← WHO Malaria Treatment Guidelines ↓ Retrieved context injected into Gemma 4 prompt ↓ Response in CHW's language + source citations ``` --- ## Quick Start ### 1. Install dependencies ```bash pip install -r requirements.txt ``` ### 2. Install and start Ollama ```bash # Install Ollama: ollama.ai ollama pull gemma3:4b # lightweight (4B params — good for demo) ollama pull gemma3:27b # higher quality (requires ~16GB RAM) ollama serve ``` ### 3. Ingest clinical guidelines ```bash # Downloads WHO/MSF PDFs and builds the ChromaDB vector store python src/ingest.py --download ``` ### 4. Launch Daktari AI ```bash python src/app.py # opens at localhost python src/app.py --share # generates a public Gradio link for demos ``` --- ## Running on K …