salmonswan/mal-ai-flutter

On-Device MedGemma for Frugal Diagnosis of Malaria and Cervical Cancer in Sub-Saharan Africa. # MAL-AI — Frugal AI Diagnostic Platform for Africa A Flutter app for the MedGemma Impact Challenge 2026. Three diagnostic modules -- malaria PPG triage, cervical cancer VIA screening, and blood smear microscopy -- powered by MedGemma 4B multimodal reasoning across three deployment tiers. ## Getting Started ### Prerequisites - Flutter SDK >= 3.10 - Android device or emulator (API 21+) - ~3 GB free storage for on-device models (optional) ### Build & Run ```bash git clone && cd mal-ai-flutter flutter pub get flutter run # debug on connected device flutter build apk # release APK (~131 MB) ``` ### On-Device AI Setup (Tier 1) The app defaults to on-device inference. From the app: 1. Open **Settings** (gear icon) > **Manage Models** 2. The screen shows your device RAM and recommended quantization 3. Tap **Download** on a MedGemma GGUF model (Q4_K_M recommended for most phones) 4. Download the **Multimodal Projector** (mmproj-F16.gguf, ~812 MB) for image analysis Models are downloaded from HuggingFace and stored in the app's private directory. ### Alternative: Local Hub (Tier 2) Run MedGemma on a nearby laptop via Ollama: ```bash ollama run hf.co ``` In the app, switch to **Local Hub** in Settings and set the URL to `http:// :11434/v1`. ### Alternative: Cloud (Tier 3) Configure Vertex AI credentials in Settings for cloud-hosted MedGemma 27B. ## Architecture ``` Tier 1: On-Device Tier 2: Local Hub Tier 3: Cloud llamadart (llama.cpp) --> Ollama on LAN laptop --> Vertex AI MedGemma 4B + mmproj OpenAI-compatible API Google Cloud Adaptive GGUF quant 16 GB RAM sufficient Full precision ``` The app auto-selects the best available tier, or the user can choose in Settings. ## Diagnostic Modules ### Malaria Triage ``` Home > Intro > PPG Capture (30s) > Symptoms > MedGemma > Result ``` - Camera-based PPG extracts HR, HRV, estimated hemoglobin, waveform morp …