Building UbuntuAfya - an offline-first mobile AI assistant for Community Health Workers in rural Kenya

I built UbuntuAfya — an offline-first mobile AI assistant for Community Health Workers in rural Kenya, where the main feature is a GGUF LLM running directly on a $150 Android phone that CHWs can talk to about patient symptoms — in Kiswahili (still revamping) or English — with zero internet.

Demo context:

The honest builder story is: we didn’t start from scratch — we started from the right place.

We found PocketAI, an open source Flutter app that runs GGUF models on-device via llama_flutter_android (llama.cpp on ARM64 Android). The architecture was exactly what we needed — model download, storage management, streaming inference, chat UI. We adapted it and asked: what if this was a medical tool for someone walking 10km through rural Kenya with no signal?

Here’s what we built on top of it — and the demo shows all three:

1. The offline AI symptom checker — the big one We swapped in MedGemma — Google’s medical-domain Gemma variant, quantized to GGUF. A CHW describes patient symptoms and MedGemma streams a structured response directly on the Android chip — what it could be, what to do, when to refer. No API key. No server. No signal. I’ll show the model loading, the streaming tokens, and the resumable download system I built so a CHW on patchy 3G doesn’t lose a download when signal drops.

2. The AI patient history report When connectivity returns, one tap on any patient’s visit timeline sends their full history to openai/gpt-4o-mini via OpenRouter. Back comes a structured plain-text medical analysis — probable conditions, recommendations, warning signs — formatted to print directly from the phone.

3. Speech_to_Text - lets CHWs describe symptoms by voice. flutter_tts reads AI responses back aloud. The system detects language automatically and responds in kind. The medical safety guardrail — no diagnosis, no prescriptions — is written in both languages inside the system prompt, so it can’t be bypassed by switching languages.

The interesting builder question the demo answers: how do you take a general-purpose open source LLM tool and responsibly adapt it for healthcare in a low-resource, multilingual, offline field setting?