Amalest/africa-shield-ai

# Africa Shield AI **Big vision:** Africa Shield AI predicts natural disasters (floods, droughts, heatwaves, wildfires, cyclones, earthquakes) across Africa and sends early warnings to at-risk communities before disasters happen, prioritizing accessibility for people without smartphones or reliable internet. **Scoped hackathon demo:** "Last-Mile Alert AI" — flooding only, working end-to-end: a rules-based flood risk score shown alongside a genuine trained ML model as a second opinion, an API that turns the result into a human-readable alert, translation into a local African language, and a simulated SMS/USSD/WhatsApp send, all shown on a simple web dashboard. **Future:** expand to the other hazards, train the ML model on real historical data instead of synthetic data, a real SMS gateway, and an offline-first mobile app (see Future Improvements below). Built for the **"AI for All Hackathon: Building Inclusive Solutions for Early Warning and Disaster Resilience,"** organized by the African Youth Advisory Board on Disaster Risk Reduction (AYAB-DRR) under the African Union. ## Status: backend logic real, frontend in progress (updated 2026-08-10) - **Backend risk scoring and translation are real**, not stubbed. `POST /api/risk-check` and `GET /api/regions` compute live from the rules-based model in `backend/app/models/risk_model.py` and the hardcoded translation dictionary in `backend/app/models/translations.py`. See `docs/progress-log.md` for thresholds, assumptions, and what's still unverified. - **A second, genuinely trained ML model now runs alongside the rules-based one** (`ml_risk_level`/`ml_risk_score`, additive fields on both endpoints) — a `scikit-learn` logistic regression, trained on synthetic data standing in for real historical flood data. See `docs/architecture.md`'s "Two risk scores, on purpose" section for why both are kept. - `GET /api/alerts` is **still an intentional simulated stub** — no real SMS/USSD gateway this week (documented future improveme …