Farming 🌱 — Local Crop Advisor

Upload a photo of any crop → get a diagnosis, immediate actions, and contextual advice powered by a local vision model. No cloud, no API costs, no connectivity required.

How it works

Photo + Crop + Location
        │
  ┌─────▼──────┐
  │ Photo      │  llama3.2-vision  →  visual diagnosis
  │ Module     │
  └─────┬──────┘
        │
  ┌─────▼──────┐
  │ Weather    │  OpenWeatherMap (cached, offline fallback)
  │ Module     │
  └─────┬──────┘
        │
  ┌─────▼──────┐
  │ Inference  │  llama3 (text)  →  structured action plan
  │ Engine     │
  └─────┬──────┘
        │
  Diagnosis  ·  Severity  ·  Actions  ·  Timeline

Adding a new parameter (soil, water, market prices) = one new file in modules/.

Setup on Linux (Ubuntu 22/24)

1. Install Ollama

curl -fsSL https://ollama.com/install.sh | sh

2. Pull the models

# Vision model — 8 GB download, needs 12 GB RAM
ollama pull llama3.2-vision:11b

# Text inference model — 4.7 GB (use the one you already have)
ollama pull llama3:latest

Lower-RAM option: replace llama3.2-vision:11b with llava:7b (4.7 GB, needs 8 GB RAM).
Set VISION_MODEL=llava:7b in your .env.

3. Clone and install

git clone https://github.com/gkalidas/farming.git
cd farming
python3 -m venv venv && source venv/bin/activate
pip install -r requirements.txt

4. Configure (optional)

cp .env.example .env
# Edit .env — add OPENWEATHER_KEY for live weather, leave blank for offline

5. Run

uvicorn main:app --host 0.0.0.0 --port 5002
# Open http://localhost:5002 — or use your laptop's IP on mobile

Supported crops (Phase 1)

• Pomegranate
• Tomato
• Grape
• Onion
• Cotton

Adding a new parameter module

Create modules/soil.py:

from modules.base import BaseModule, ModuleContext

class SoilModule(BaseModule):
    name = "soil"

    async def get_context(self, crop: str, location: str, date: str) -> ModuleContext:
        # fetch or read soil data
        return ModuleContext(
            module_name=self.name,
            available=True,
            summary="Soil pH 6.2, nitrogen low, potassium adequate",
            detail={"ph": 6.2, "nitrogen": "low"},
        )

Register it in main.py:

from modules.soil import SoilModule
registry.register(SoilModule())

That's it. The inference engine picks it up automatically.

Stack

Backend:   FastAPI + SQLite (WAL)
Frontend:  Vanilla JS, mobile-first, no build step
Models:    Ollama — llama3.2-vision (vision) + llama3 (inference)