PrithviPulse

Inspiration

India has over 150 million farmers, yet crop diseases silently wipe out 30–40% of yields every year, costing billions in lost income. Most small farmers don’t have access to agricultural experts. Instead, they rely on guesswork, word of mouth, or outdated practices. By the time disease symptoms are clearly visible, the damage is often already done.

I kept asking myself: What if every farmer had an agricultural expert in their pocket? What if taking a photo of a diseased leaf could instantly tell them what’s wrong, how serious it is, and exactly what to do next?

That question led to PrithviPulse.

Built using Google’s Gemini 3 Vision API and agricultural domain knowledge, PrithviPulse aims to make crop disease detection fast, accurate, and accessible. The name “Prithvi” (Sanskrit for Earth) reflects the core idea behind the project: healing the Earth, one crop at a time.

What I Learned

Exploring Gemini 3 Vision’s Multimodal Strength

Learned how to design precise prompts for real-world agricultural image analysis
Discovered that Gemini 3 doesn’t just identify diseases—it explains them and recommends treatments
Learned to enforce structured JSON outputs for reliable UI rendering
Used safety settings to ensure responsible and consistent agricultural advice

Building a Hybrid AI System

Designed a fail-safe architecture: Gemini 3 as the primary model with a local CNN as fallback
Ensured 100% uptime, even when cloud APIs hit rate limits
Learned how to balance cloud intelligence with offline reliability

Why Gemini 3 Works So Well for Agriculture

Excels at identifying subtle visual differences (e.g., early vs late blight)
Goes beyond classification to provide actionable treatment steps
Produces clear, farmer-friendly explanations
Handles messy, real-world images with dirt, shadows, and damaged leaves

Key Insight: Traditional CNNs stop at labels. Gemini 3 delivers end-to-end agricultural guidance, which makes it far more practical for farmers.

How I Built It

System Architecture

React Frontend (TypeScript + Vite + Tailwind)
    ↓ HTTP/REST API
FastAPI Backend (Python)
    ↓
Hybrid AI Pipeline:
1. Gemini 3 Vision (Primary)
   - Leaf image analysis
   - Structured JSON output
   - Treatment recommendations
2. Local CNN Model (Fallback)
   - TensorFlow/Keras
   - 38 disease classes
   - Fully offline

Technology Stack

Frontend

React 18.3.1 + TypeScript – Type-safe, scalable UI
Vite 6.4.1 – Fast development and builds
Tailwind CSS – Mobile-first, responsive design
Web Speech API – Text-to-speech for treatment steps

Backend

Python + FastAPI – Async REST API
Google Gemini 3 (gemini-3-flash-preview) – Vision AI
TensorFlow + Keras – Local CNN inference
Uvicorn – Production-ready ASGI server

AI Models

Primary: Gemini 3 Vision

Input: Leaf images
Output: Structured diagnosis + detailed treatment

Fallback: Custom CNN (plant_disease_model.h5)

Trained on 87,000+ images
Covers 38 disease classes
Works completely offline

Core Implementation

model = genai.GenerativeModel("gemini-3-flash-preview")

response = model.generate_content([
    genai.upload_file(image_path),
    """Analyze this leaf image and provide:
    1. Disease name and confidence
    2. Professional summary
    3. Step-by-step treatment (physical + chemical)
    4. Preventive measures
    Return the response as structured JSON."""
])

Unique Features

Visual Treatment Timeline Step-by-step treatment flow with urgency indicators, progress tracking, and text-to-speech
Chemical Safety System Automatically detects chemical treatments, displays ⚠️ warnings, and provides dosage guidance
Offline-First Design Works without internet using the local CNN model
Mobile-Optimized UI Designed for field use on devices from 320px to 4K screens

Challenges I Faced

Gemini API Quota Limits

Problem: API quota was exhausted during repeated testing. Solution: Added automatic fallback to the local CNN model with retry logic and quota monitoring. Result: 100% uptime, even when Gemini is unavailable ✅

Structured Output from Vision Models

Problem: Gemini naturally returns free-form text, but the UI required consistent JSON. Solution: Carefully engineered prompts with strict JSON schema enforcement and fallback parsing. Result: 95% valid JSON responses ✅

Real-World Image Noise

Problem: Farm images often have poor lighting, dirt, shadows, and damaged leaves. Solution: Relied on Gemini 3’s strong visual reasoning and added preprocessing with confidence thresholds. Discovery: Gemini 3 significantly outperforms traditional CNNs on real farm photos ⭐

Agricultural Accuracy

Problem: Many diseases share very similar visual symptoms. Solution: Used domain-specific prompting and Gemini’s multimodal reasoning. Result: 88% accuracy, outperforming the local CNN by 12% ✅

Large Model File Size

Problem: GitHub rejected the 44MB H5 model file. Solution: Configured Git LFS and optimized repository setup. Lesson: Always plan for large ML assets early.