AI-Powered Crop Diagnosis, Advisory & Multilingual Assistance

An end-to-end, mobile-first platform that helps farmers diagnose crop diseases, get intelligent recommendations, and interact using voice or chat — even in low-connectivity environments.

---

🚀 Live Demo

👉 https://crop-ai-one.vercel.app/

---

🚀 Overview

Crop AI is a full-stack, AI-driven agriculture platform designed to assist farmers with:

• 🌱 Plant disease detection from leaf images
• 💬 Conversational AI chat assistant
• 🌾 Crop recommendation & yield prediction
• 🌍 Region-aware agricultural guidance
• 🗣 Multilingual + Speech-based interaction
• 📱 Offline-friendly mobile experience

The project is built as a TurboRepo monorepo, combining mobile, web, backend, and ML services into a single, scalable codebase.

---

🧠 System Architecture

            ┌──────────────┐
            │  Mobile App  │  (React Native + Expo)
            └──────┬───────┘
                   │
        ┌──────────▼─────────┐
        │   Backend Server   │  (Node.js + Express)
        │  Auth • Chat • API │
        └──────┬───────┬─────┘
               │       │
     ┌─────────▼───┐   ▼
     │  ML Server  │  MongoDB
     │  (FastAPI)  │  (Mongoose)
     └─────────────┘

---

✨ Features Status

🤖 AI Chat Server

• Conversational assistant powered by Ollama
• Custom tool calling system (not LangChain-based)
• Supports reasoning, structured responses, and agriculture-focused queries
• Designed to later plug into LangChain.js if needed

🌿 Plant Leaf Disease Detection

• Dedicated ML server (FastAPI)
• Image-based disease detection
• Model evolution:
  • ResNet9 → ResNet18 → ResNet50 (current)
• Optimized for mobile image inputs

🌾 Crop Recommendation & Prediction

• Multiple classical ML models trained and evaluated

🗣 Voice & Language Support

• Speech-to-Text and Text-to-Speech using (On Device using APIopenai/whisper)
• Multilingual support for farmer accessibility

🔐 Authentication

• Clerk for secure, scalable authentication
• Works across mobile

---

🛠 Tech Stack

Frontend

• React Native + Expo
• React + Vite (Web)
• Clerk Authentication

Backend

• Node.js + Express
• MongoDB + Mongoose
• Ollama (LLM runtime)

Machine Learning

• FastAPI
• PyTorch
• Scikit-learn
• XGBoost

DevOps

• TurboRepo
• Docker & Docker Compose
• pnpm

---

Getting Started

1. Install Dependencies

Install all project dependencies:

pnpm install

If you face issues with Expo Go:

cd apps/native
pnpm install

Set up the ML server:

cd apps/ml-server
uv venv
uv sync

Optional setups:

cd apps/server
pnpm install

cd apps/web
pnpm install

---

2. Database Setup

This project uses MongoDB with Mongoose.

1. Ensure MongoDB is installed and running.
2. Update apps/server/.env with your MongoDB connection URI.

Get API Keys from

• WEATHER API
• TAVILY
• CLOUDINARY
• CLERK

---

3. Running the Project

Start the development servers:

pnpm dev

• The mobile app can be run via Expo Go.
• Backend API is available at: http://localhost:3000.

---

Project Structure

crop-ai/
├── apps/
│   ├── native/      # Mobile app (React Native + Expo)
│   ├── ml-server/   # Machine Learning server (FastAPI)
│   ├── server/      # Backend API (Express)
│   └── web/         # Web app (React + Vite)

---

Available Scripts

Script	Description
pnpm dev	Start all applications in development mode
pnpm build	Build all applications
pnpm dev:native	Start the React Native/Expo development server
pnpm dev:web	Start only the web application
pnpm dev:ml-server	Start only the ML FastAPI server
pnpm dev:server	Start only the backend server
pnpm check-types	Check TypeScript types across all apps

---

Features

• AI-powered disease detection from leaf images
• Crop Recommendation (XGBoost)
• Crop Yield Prediction (XGBoost)
• RealTime Mandi Market Analysis
• Speech to Text and vice-versa supported
• Multilingual chat assistant for farmers
• Crop and rotation guidance tailored to local regions
• Works offline and on mobile devices
• Easy-to-use mobile applications

🧪 Machine Learning Models

🌿 Plant Disease Detection

Image-based classification

• Backbone: ResNet50
• Input: Leaf images
• Output: Disease class + confidence

Available Models

• Disease Recommendation Models: 3
• Evolution: ResNet9 → ResNet18 → ResNet50