yusufpy/Build-with-Gemma-AI-for-Africa-Hackathon-Minna-2026

# Ilera — Field Triage Assistant **Multimodal, multilingual, agentic triage support for community health workers — built on Gemma 4.** > Built for the *Build with Gemma: AI for Africa* Hackathon (Minna, 2026) 🔗 **Live demo: ilera-mc67.onrender.com A health worker describes symptoms (and optionally attaches a photo). Gemma returns a structured triage assessment — severity, likely conditions, red flags, recommended action — which can be output in English, Hausa, or Yoruba. The system then autonomously drafts a referral note and schedules a follow-up based on urgency ## Why this isn't "just a chat prompt" - **Multimodal reasoning, not captioning.** Gemma 4 reasons jointly over a photo and free-text notes to produce a differential and flag risks a non-specialist might miss (e.g. the elevated danger of facial infections spreading via the "danger triangle"). - **Structured output that drives real logic.** Gemma's response is forced into a strict JSON schema via the API's `responseSchema` feature. That object — not free text — is what the referral and follow-up scheduling logic actually runs on. - **Local language, done carefully.** Clinical reasoning stays in English internally (more reliable to validate). Only the final output is translated, and it's back-translated and confidence-checked before being shown — Gemma auditing Gemma's own translation. See `translate.py`. - **Agentic follow-through.** The triage result triggers real downstream action: a referral note is drafted and a follow-up is scheduled automatically based on urgency. See `agent.py`. ## Setup To run it locally: ```bash cd ilera python3 -m venv venv source venv/bin/activate # Windows: venv\Scripts\activate pip install -r requirements.txt cp .env.example .env # edit .env and paste in your ownGemma API key (from Google AI Studio) ``` ### Find your exact model name first — don't guess it Gemma 4 uses capability-tier naming (e.g. `gemma-4-31b-it`, `gemma-4-26b-a4b-it`), not the p …