AgriPulse

Farmers lose $5B yearly to crop diseases. AgriPulse uses Gemini AI to diagnose plants from photos, predict yields, and speak advice in 6 Indian languages—bridging tech to the last-mile farmer.

Comment

Gemini 3 Integration: AgriPulse

AgriPulse is fundamentally built on Gemini 3's advanced multimodal capabilities, making AI-powered agricultural guidance accessible to small-scale farmers. The application leverages Gemini 3 as its core intelligence engine in three critical ways:

1. Multimodal Crop Analysis (Gemini 3.0 Flash)
Farmers capture crop photos and describe symptoms via voice. AgriPulse combines these inputs—image + voice transcription + weather context—into a unified Gemini analysis prompt. Gemini 3.0's Vision API identifies diseases/pests with 94%+ accuracy, understanding subtle visual cues like leaf discoloration, pest damage, and fungal patterns. This multimodal approach surpasses single-input analysis, providing holistic agricultural insights.

2. Reasoning-Based Yield Prediction (Gemini 3.0 Pro)
The yield prediction engine uses Gemini 3.0 Pro's advanced reasoning capabilities to synthesize complex, interconnected variables: historical yields, weather forecasts, soil conditions, applied treatments, and climate patterns. Gemini 3.0's reasoning mode breaks down its logic transparently, showing farmers exactly how each factor impacts their expected yield—building trust through explainability.

3. Language-Adaptive Guidance
Gemini understands agricultural terminology across 6 Indian languages (Tamil, Telugu, Kannada, Malayalam, Hindi, English), enabling farmers to input problems in their native language and receive recommendations back in the same language. This eliminates language barriers—the core blocker for agricultural tech adoption in India.

Without Gemini 3's multimodal strength, reasoning capabilities, and language support, AgriPulse would be significantly limited. Gemini 3 is not a feature—it's the foundational intelligence that makes this application possible.

Inspiration

My father has been farming for over 50 years. Growing up, I watched him predict weather patterns, identify diseases by looking at leaves, and estimate yields with remarkable accuracy—all from decades of experience.

But over the past 10 years, everything changed. Climate patterns became unpredictable. Diseases he'd never seen before started appearing. His 50 years of wisdom—built on a climate that no longer exists—couldn't keep up. Every farmer in my village faces the same struggle.

AgriPulse was born from watching my father struggle and wanting to give him—and millions of farmers like him—a fighting chance against an unpredictable future.

What it does

AgriPulse is an AI-powered farming assistant that helps Indian farmers:

  • 🔬 Crop Doctor - Snap a photo of a plant, describe symptoms in your language(voice/text), get instant disease diagnosis with treatment recommendations.
  • 📊 Yield Predictor - Combines weather forecasts, soil data, and crop information to predict harvest yields.
  • 🗣️ Multilingual Voice - Supports 6 Indian languages (English, Hindi, Tamil, Kannada, Telugu, Malayalam) with voice input/output.
  • 📱 Offline-First - Access past diagnoses without internet connectivity (to assist village farmers living in low Internet connectivity region).

How we built it

  • Frontend: Flutter Web with Riverpod (state management) and GoRouter (navigation).
  • Backend: FastAPI (Python) for API orchestration.
  • AI Engine: Google Gemini 3.0 Flash for vision-based diagnosis, multilingual text generation, and structured JSON outputs.
  • Storage: Hive (offline-first local storage) + Firestore (cloud sync).
  • Speech: Google Cloud Text-to-Speech for voice advice playback.
  • Weather: OpenWeatherMap API for real-time 14-day forecasts.
  • Development: Built with assistance from Antigravity (Google's AI coding assistant).

Challenges we ran into

Building the Feedback Loop - The biggest challenge wasn't code—it was designing how farmers can correct AI mistakes. Unlike tech-savvy users who can articulate "the model got this wrong," farmers simply know "this advice didn't work." We built a "Challenge Diagnosis" feature that captures disagreements and reasoning, creating a feedback loop to improve future predictions.

Translating Farmer Knowledge to AI Prompts - Farmers describe problems differently than textbooks. "Leaves are looking tired" means something specific to them. We had to bridge the gap between colloquial descriptions and structured AI queries while preserving the farmer's original context.

Designing for Users Who Can't Test - My father can't file bug reports. Every assumption about "obvious" UI flows had to be questioned. Voice-first wasn't a feature—it was survival for users with low literacy.

Accomplishments that we're proud of

  • Working E2E Crop-Diagnosis - From photo capture to AI diagnosis to voice playback and get feedback with additional factors from Farmers (if inconsistencies with the AI diagnosis).
  • Working E2E Yield-Prediction - Weather-aware harvest forecasting with 7-day predictions and confidence scores.
  • ✅ 6 Language Support - Full localization.
  • ✅ Voice-First Interface - Speech-to-text input + Text-to-speech output for accessibility.
  • ✅ Offline Capability - Historical records persist locally using Hive.
  • ✅ Beautiful, Accessible UI - Context-aware icons, colors, and labels for healthy vs diseased plants.
  • ✅ Solo Project - Built end-to-end by a single developer in hackathon timeframe (learned about the hackathon 13 days before the deadline and able to complete the milestone within 10 days).

What we learned

  • Gemini 3.0 Flash is incredibly powerful for multimodal tasks—vision + text + multilingual in a single prompt.
  • Defensive programming is essential with AI APIs—always expect the unexpected.
  • Accessibility (voice, local languages) isn't optional—it's the difference between adoption and abandonment.
  • AI-assisted development (Antigravity) dramatically accelerates prototyping while maintaining code quality.
  • Technology only matters if it reaches people who need it. Simplicity is a feature, not a compromise.

What's next for AgriPulse

  • 🌾 Crop-Specific Models - Fine-tuned diagnosis for regional crops (onion, rice, cotton).
  • 🛰️ Satellite Integration - NDVI/vegetation indices for field-level health monitoring.
  • 🏪 Market Price Alerts - Real-time mandi prices with sell recommendations.
  • 👨‍🌾 Expert Connect - Video consultations with agricultural officers.
  • 📴 Full Offline Mode - Edge AI using TensorFlow Lite for zero-connectivity areas.
  • 🤝 Farmer Network - Community feature to share local pest/disease warnings and Government Schemes.

Built with ❤️ for Global Farmers—and for my Father.

🧪 Testing the Demo

For Judges & Evaluators:
I've provided sample crop disease images to test the AI Crop Doctor feature:

Sample Images:

How to Test:

  • Visit https://agri-pulse-firebase.web.app
  • Click "Skip Login (Demo)" on the login screen
  • Navigate to "Crop Doctor" from the dashboard
  • Upload one of the sample images from the links above
  • Describe symptoms via voice or text (optional)
  • View the AI-generated diagnosis and treatment plan in seconds

Built With

Share this project:

Updates