Uthmannabeel/shuka

Offline agronomy assistant for the hardware Africa actually has — ADTC 2026 (Agriculture track) # Shuka — offline agronomy assistant **Website: uthmannabeel.github.io **Africa Deep Tech Challenge 2026 · Agriculture track.** *Shuka* is Hausa for "to plant / to sow." Nigeria's agricultural extension system is stretched to roughly one extension worker per several thousand farmers, and the obvious fallback — asking an AI assistant — fails exactly where farming happens: rural areas with weak or unaffordable connectivity, on low-end hardware. Shuka is an agronomy assistant that runs **entirely offline** on the hardware cooperatives and agro-dealers actually own (the ADTC Standard Laptop: 8GB RAM, integrated graphics, i5-class CPU). One offline laptop at a cooperative serves a whole community. Small language models make fluent, confident, and **dangerously wrong** agronomists — our baseline runs had a bare 1B model recommending growing cassava from seed and inventing fertiliser arithmetic. Shuka's answer is architectural, not cosmetic: the model supplies *language*, a curated corpus of real extension literature supplies the *facts*. ``` question → on-device embedding (MiniLM, 384d, quantized ONNX) → exact cosine search over indexed extension literature (FAO / IITA / CABI-ASHC / IRRI — ~850 pages, license-verified) → top-4 passages + question → Llama 3.2 1B Instruct Q4_K_M (llama.cpp) → answer constrained to sources, with page-level citations ``` **Guardrail:** if no passage clears a similarity floor, the model is not called at all — Shuka says it doesn't have sources and points the farmer to their local extension office. Silence beats confident error when the cost of being wrong is somebody's growing season. ## Quickstart ```bash bash install.sh # checks Node 20+, installs deps, fetches the model (~0.8 GB), pre-caches npm run serve # web app at localhost ``` After `install.sh` completes, no network is used at all. The web app presents each answer as a **manual-style answer sheet**: the text cites its sources like a …