Offline-first multilingual AI crop disease advisor for West African smallholder
# PokouAI 🌱
**Offline-first, multilingual AI crop disease advisor for West African smallholder farmers.**
Built for the Gemma 4 Good Hackathon — submission deadline May 18, 2026.
- 📱 **100% offline** inference on entry-level Android phones (2–3 GB RAM)
- 🗣️ **4 languages**: French, Dioula, Baoulé, English
- 🧠 **Gemma 4 E2B** (multimodal) fine-tuned on cocoa diseases with QLoRA + Unsloth — ~1.5 GB GGUF fits low-end devices. E4B variant built as optional premium for ≥4 GB phones.
- 🌾 **Primary use case**: cocoa disease diagnosis in Ivory Coast (5M+ farmers, mostly offline)
---
## The problem
Ivorian cocoa farmers lose 30–40% of their harvest to preventable diseases. Extension agents are scarce, internet is unreliable, and most agronomy apps assume fluent French literacy and 4G. A farmer holding a diseased pod today has no fast way to identify *black pod rot* from *frosty pod rot* — and the treatments are different.
PokouAI runs entirely on the phone: point camera at pod, get a disease name, treatment steps, and prevention advice in the farmer's own language. Airplane mode works. A $50 Android works.
---
## Quick start
> 📖 **If you're a reviewer running this for the first time, read docs/STORY.md first.** It owns the constraints (Mac RAM, training-subset choices, Unsloth fine-tune) and where this is heading (Fair Trade Côte d'Ivoire partnership, first cooperative pilot next month).
### First run — 60 seconds to the Farmer Agent demo
```bash
# Prereqs: Node 20 (via nvm), Xcode + iOS sim (or Android Studio + emulator), pnpm
cd app/
pnpm install # if you see deps land in the workspace root, see Troubleshooting
pnpm typecheck # must pass clean before reporting bugs
pnpm start --clear
pnpm ios # or: pnpm android
```
In the app, on the Home screen, tap the purple **🎬 Farmer Agent — Skill demo** tile. The screen shows the same disease (black pod) and same image at **5 flow stages × 3 skill levels** — t …