Zuku03/varimi-sentinel

Offline edge-AI crop-climate & market advisory for Zimbabwe - POTRAZ AI4I Track 3 # VaRimi Sentinel Offline-first **edge-AI crop-climate risk & market advisory** for Zimbabwe. POTRAZ AI4I Grand Challenge — **Track 3 (Development)**. For a given **district + crop + month**, VaRimi Sentinel predicts: 1. **Climate-crop risk band** (Low / Medium / High) with driver explanation 2. **Yield outlook** (tonnes per hectare) 3. **Farmgate price direction** (down / flat / up) …and renders a plain-language recommended action for an agricultural extension officer, delivered on a low-end Android device or via USSD/SMS in Shona, Ndebele or English. ## Why AI (not a rule table) Risk, yield and price are joint non-linear functions of rainfall, vegetation (NDVI), pest pressure, irrigation coverage and input availability, conditioned on season and agro-ecology. A static rule/SQL table cannot generalise across unseen district × crop × month combinations. See `docs/AI_JUSTIFICATION.md` (Phase 3). ## Data Trained on `Datasets/02_agriculture_climate_market_signals.csv` — a **synthetic aggregate** sample from the official AI4I Design-Track pack (not official statistics). Augmented with a fully-disclosed Gaussian-copula synthesizer, validated with statistical correlation tests. See `docs/DATASET_STATEMENT.md`. ## Layout ``` src/varimi/{data,features,synth,model,edge,serving} # library code tests/ # pytest suite reports/ # validation + benchmark reports docs/ # dataset statement, model card, AI justification proposal/ # 10-page Track-3 proposal (Phase 5) ``` ## Quickstart ```bash python -m venv .venv .venv\Scripts\python -m pip install -e ".[dev,demo]" pytest ``` ## Status Prototype / MVP under active build for the AI4I Grand Challenge. Synthetic data only; not for operational policy decisions. ## Reproduce end-to-end ```bash python -m varimi.model.train # train + reports/m …