AGROSAGE - Smart Soil Monitoring for Modern Farmers

A hardware and software project get soil data, works offline, provides AI insights, and supports multiple languages.

Comment

Inspiration

The inspiration for AgroSage stems from a critical challenge facing millions of farmers worldwide: the lack of accessible, affordable, and reliable soil health monitoring tools. Traditional soil testing methods are expensive, time-consuming, and often require sending samples to distant laboratories, making them impractical for small-scale and resource-constrained farmers.

Key Problems We Aimed to Solve:

  1. Limited Internet Connectivity: Many agricultural regions, especially in developing countries, have poor or no internet connectivity. Farmers need tools that work offline.

  2. Language Barriers: Agricultural knowledge and technology are often only available in English, creating barriers for farmers who speak regional languages like Hindi and Bengali.

  3. High Costs: Professional soil testing services are expensive and not accessible to small-scale farmers who need frequent monitoring.

  4. Delayed Results: Traditional lab testing takes days or weeks, while farmers need immediate insights to make timely decisions about irrigation, fertilization, and crop management.

  5. Privacy Concerns: Farmers are hesitant to share sensitive agricultural data with cloud services, preferring local processing.

  6. Technical Complexity: Existing solutions require technical expertise, making them inaccessible to farmers with limited technical knowledge.

Our Vision: Create an affordable, offline-first, multi-language mobile application that empowers farmers to monitor soil health in real-time using Bluetooth-enabled sensors, with AI-powered insights that work entirely on-device for privacy and speed.

What it does

AgroSage is a comprehensive mobile application that transforms how farmers monitor and manage soil health. Here's what it does:

Core Functionality

  1. Real-Time Soil Monitoring

    • Connects to ESP32-based soil probes via Bluetooth Low Energy (BLE)
    • Captures 10 critical soil parameters in real-time:
      • Soil moisture and temperature
      • pH level and Electrical Conductivity (EC)
      • NPK values (Nitrogen, Phosphorus, Potassium)
      • Air temperature and humidity
      • Light intensity
    • Displays data in an intuitive, easy-to-understand format

  2. AI-Powered Insights

    • On-Device AI (iOS): Uses ARM Neural Engine to generate instant insights (~2ms inference time)
    • Cloud AI (All Platforms): Server-side AI processing when internet is available
    • Rule-Based Fallback: Offline insights when AI services are unavailable
    • Provides actionable recommendations for:
      • Fertilizer application (NPK supplementation)
      • Irrigation scheduling
      • pH management
      • Salinity control
      • Environmental optimization

  3. Offline-First Design

    • All core features work without internet connectivity
    • Data stored locally on the device
    • AI insights generated on-device (iOS) or via rule-based system
    • Cloud sync available when internet is accessible

  4. Multi-Language Support

    • Complete app translation in English, Hindi, and Bengali
    • All AI insights fully translated
    • Dynamic language switching with instant updates
    • Offline translation (no internet required)

  5. Data Management

    • Save and review historical soil measurements
    • Automatic GPS location capture
    • Cloud backup (optional)
    • Export and share capabilities

  6. User-Friendly Interface

    • Modern, intuitive UI designed for farmers
    • Real-time status indicators
    • Color-coded health metrics
    • Clear, actionable recommendations

Key Differentiators

  • Privacy-First: On-device AI processing ensures farmer data never leaves their device
  • Instant Results: ARM Neural Engine provides insights in milliseconds
  • Accessibility: Works in remote areas without internet
  • Affordability: Uses low-cost ESP32 hardware instead of expensive lab equipment
  • Multilingual: Breaks language barriers for regional farmers

How we built it

Phase 1: Foundation & Architecture

Technology Stack Selection:

  • React Native + Expo: Chose for cross-platform development and rapid iteration
  • TypeScript: Ensured type safety and better code maintainability
  • Expo Router: Implemented file-based routing for navigation

Project Structure:

  • Modular architecture with separate contexts for BLE, Readings, and Language
  • Component-based UI with reusable elements
  • Service layer for API calls and AI processing

Phase 2: Bluetooth Integration

BLE Implementation:

  • Integrated react-native-ble-plx for BLE communication
  • Implemented device scanning with automatic discovery
  • Created robust connection management with error handling
  • Handled fragmented JSON payloads (BLE MTU limitations)
  • Optimized MTU to 185 bytes for efficient data transfer

Key Challenges Solved:

  • Fragmented data reconstruction from multiple BLE packets
  • Connection stability across different Android/iOS versions
  • Proper cleanup of BLE subscriptions to prevent crashes

Phase 3: AI Model Development

Model Training:

  • Designed neural network architecture: 10 inputs → 64 hidden → 64 hidden → 32 hidden → 40 outputs
  • Generated 10,000 synthetic training samples based on optimal soil parameter ranges
  • Trained PyTorch model for 100 epochs, achieving final loss of 0.1180
  • Total model parameters: 8,264

ExecuTorch Export:

  • Converted PyTorch model to ExecuTorch format (.pte)
  • Implemented multiple export strategies (CoreML, Portable, XNNPACK)
  • Successfully exported using portable backend
  • Model size: ~42.5 KB

Integration:

  • Integrated react-native-executorch for on-device inference
  • Configured Metro bundler to handle .pte files
  • Implemented asset loading and model caching
  • Created output parsing logic to convert 40-dimensional vectors to structured insights

Phase 4: Multi-Language System

Translation Infrastructure:

  • Created offline translation dictionary with 200+ keys
  • Implemented language context with AsyncStorage persistence
  • Built translation function with variable substitution support
  • Added dynamic language switching with instant UI updates

Translation Coverage:

  • All UI elements and labels
  • AI insight titles and descriptions
  • Priority labels (High, Medium, Low)
  • Reasoning messages (WHY section)
  • Summary messages
  • Error messages and alerts

Phase 5: Cloud Integration

Supabase Setup:

  • Configured Supabase Edge Functions for server-side AI processing
  • Implemented API client with timeout handling
  • Created fallback system for offline scenarios
  • Added error handling and retry logic

Phase 6: UI/UX Development

Design Implementation:

  • Created modern, clean interface with safe area handling
  • Implemented real-time status indicators
  • Built metric cards with color-coded health indicators
  • Added loading states and user feedback
  • Designed welcome splash screen
  • Created reading history and detail views

Phase 7: Testing & Optimization

Performance Optimization:

  • Optimized BLE data handling for fragmented packets
  • Implemented model caching to avoid reloading
  • Added proper error boundaries and fallbacks
  • Optimized re-renders with React best practices

Platform-Specific Optimizations:

  • iOS: Leveraged ARM Neural Engine for fast inference
  • Android: Implemented efficient rule-based fallback
  • Both: Optimized asset loading and memory usage

Challenges we ran into

1. BLE Data Fragmentation

Problem: BLE has MTU limitations (typically 20-23 bytes), causing large JSON payloads to be split across multiple packets. The app was receiving incomplete JSON strings, leading to parsing errors.

Solution:

  • Implemented a buffer system to accumulate fragmented data
  • Added logic to detect complete JSON objects (matching braces)
  • Handled newline-separated JSON chunks
  • Increased MTU to 185 bytes on Android for better efficiency

2. ExecuTorch Model Export

Problem: Initial attempts to export PyTorch model to ExecuTorch format failed due to:

  • Missing flatc compiler in PATH
  • ExecuTorch 0.3.0 compatibility issues with certain export methods
  • CoreML backend export failures

Solution:

  • Implemented multiple export strategies with fallbacks
  • Added automatic flatc path detection and PATH modification
  • Successfully exported using portable backend
  • Created comprehensive error handling and logging

3. Model Loading in React Native

Problem: react-native-executorch couldn't load the bundled .pte model file:

  • Metro bundler didn't recognize .pte files as assets
  • require() for .pte files failed
  • Asset paths were treated as URLs instead of local files

Solution:

  • Configured metro.config.js to include .pte in assetExts
  • Implemented asset loading using expo-asset and expo-file-system
  • Added fallback mechanism to copy asset to document directory
  • Created robust error handling with multiple loading strategies

4. Android Disconnect Crash

Problem: Disconnecting BLE device on Android caused NullPointerException crash:

  • subscription.remove() internally called cancelTransaction with null error code
  • Native module crashed when trying to reject promise with null

Solution:

  • Removed explicit subscription.remove() calls
  • Allowed native module to handle cleanup automatically
  • Relied on cancelDeviceConnection() to clean up subscriptions
  • Added proper state management to prevent race conditions

5. Language Translation for AI Insights

Problem: AI insights were generated once and stored as translated strings, so changing language didn't update them.

Solution:

  • Stored original reading data along with insights
  • Implemented useEffect hook to watch language changes
  • Added logic to detect ExecuTorch insights and re-translate them
  • Created translation function with variable substitution for dynamic values

6. Dependency Conflicts

Problem: npm ci failed during EAS builds due to:

  • package-lock.json out of sync with package.json
  • Missing babel-plugin-react-compiler in lock file
  • Version conflicts between expo and react-native-executorch

Solution:

  • Added babel-plugin-react-compiler to devDependencies
  • Created overrides section in package.json to force consistent versions
  • Regenerated package-lock.json with npm install
  • Ensured all transitive dependencies are resolved

7. Network Request Failures in Production

Problem: AI insights API calls failed in production builds with "Network request failed" errors.

Solution:

  • Corrected Supabase URL (removed trailing slash, used correct project URL)
  • Added proper Authorization and Apikey headers
  • Implemented timeout handling (30 seconds)
  • Added comprehensive error messages for debugging
  • Configured environment variables in eas.json

8. Model Output Interpretation

Problem: Model outputs probabilities (0.1-0.4 range) but parsing logic used 0.5 threshold, causing no insights to be generated.

Solution:

  • Lowered threshold to 0.25 for probability detection
  • Implemented sensor value validation against optimal ranges first
  • Used model outputs as confidence scores rather than condition flags
  • Added proper scaling of confidence values (20-80% range)

9. iOS App Icon Not Updating

Problem: App icon remained default even after updating app.json configuration.

Solution:

  • Regenerated native iOS project with npx expo prebuild --clean
  • Verified ios.icon path in app.json
  • Ensured icon file exists at specified path
  • Rebuilt app to include new icon assets

10. Fragmented JSON Parsing

Problem: Hardware sends JSON in chunks, but app was trying to parse incomplete JSON strings.

Solution:

  • Created resultBuffer to accumulate chunks
  • Implemented emitReadingFromJson() to detect complete JSON objects
  • Added logic to handle newline-separated JSON
  • Validated JSON before parsing to prevent errors

Accomplishments that we're proud of

1. ARM-Based On-Device AI Implementation 🎯

Successfully integrated ExecuTorch for on-device AI inference using Apple's ARM Neural Engine:

  • Performance: Achieved ~2ms inference time on iPhone 14 Pro Max
  • Privacy: All AI processing happens locally, no data sent to servers
  • Reliability: Model loads and runs consistently without errors
  • Size: Optimized model to just 42.5 KB

This is a significant technical achievement, demonstrating cutting-edge mobile AI capabilities.

2. Offline-First Architecture 🌐

Built a fully functional offline system:

  • Core features work without internet
  • AI insights generated on-device
  • Complete translation system works offline
  • Local data storage with AsyncStorage
  • Seamless cloud sync when internet is available

This makes the app accessible to farmers in remote areas with limited connectivity.

3. Comprehensive Multi-Language Support 🌍

Implemented a complete offline translation system:

  • 3 Languages: English, Hindi, Bengali
  • 200+ Translation Keys: All UI elements, insights, and messages
  • Dynamic Switching: Instant language updates without app restart
  • AI Insights Translation: All recommendations fully translated
  • Variable Substitution: Dynamic values (numbers, ranges) properly inserted

This breaks language barriers and makes the app accessible to regional farmers.

4. Robust BLE Integration 📡

Created a reliable Bluetooth communication system:

  • Handles fragmented data automatically
  • Robust connection management
  • MTU optimization for efficiency
  • Cross-platform compatibility (iOS and Android)
  • Proper error handling and recovery

The BLE implementation handles real-world scenarios like connection drops and data fragmentation seamlessly.

5. Hybrid AI Architecture 🤖

Designed a smart AI system with multiple fallbacks:

  • Primary: ARM Neural Engine (iOS) - fastest, most private
  • Secondary: Cloud AI (Supabase) - when internet available
  • Tertiary: Rule-based - always available offline

This ensures users always get insights, regardless of platform or connectivity.

6. Production-Ready Codebase 💻

Developed a maintainable, scalable codebase:

  • TypeScript: Full type safety throughout
  • Modular Architecture: Clean separation of concerns
  • Error Handling: Comprehensive error boundaries
  • Documentation: Complete README and build instructions
  • Testing: Extensive logging and debugging capabilities

7. User Experience Excellence ✨

Created an intuitive, farmer-friendly interface:

  • Modern, clean design
  • Real-time feedback and status indicators
  • Color-coded health metrics
  • Clear, actionable recommendations
  • Smooth animations and transitions

8. Cross-Platform Compatibility 📱

Successfully deployed on both iOS and Android:

  • Consistent functionality across platforms
  • Platform-specific optimizations (ARM Neural Engine on iOS)
  • Proper handling of platform differences
  • Unified codebase with minimal platform-specific code

9. Complete Documentation 📚

Created comprehensive documentation:

  • Detailed README with all features
  • Step-by-step build instructions
  • ARM-specific build guide
  • Model training documentation
  • Troubleshooting guides

10. Real-World Problem Solving 🌾

Addressed actual challenges faced by farmers:

  • Affordable soil monitoring
  • Offline functionality for remote areas
  • Multi-language support for regional users
  • Privacy-focused design
  • Instant results for timely decisions

What we learned

Technical Learnings

  1. ARM Neural Engine Capabilities

    • Learned how to leverage Apple's Neural Engine for on-device AI
    • Understood ExecuTorch export process and optimization
    • Discovered performance characteristics (~2ms inference time)
    • Gained insights into model quantization and optimization

  2. BLE Communication Challenges

    • Learned about MTU limitations and fragmentation
    • Understood the importance of proper connection lifecycle management
    • Discovered platform-specific BLE behaviors (iOS vs Android)
    • Learned to handle fragmented JSON payloads efficiently

  3. React Native Native Modules

    • Gained experience integrating complex native modules
    • Learned about Metro bundler configuration for custom assets
    • Understood asset loading strategies in React Native
    • Discovered platform-specific module behaviors

  4. Offline-First Architecture

    • Learned to design systems that work without internet
    • Understood the importance of local storage and caching
    • Discovered strategies for offline AI processing
    • Learned to balance offline and online features

  5. Internationalization Best Practices

    • Learned to implement offline translation systems
    • Understood the complexity of translating dynamic content
    • Discovered strategies for language-aware AI outputs
    • Learned to handle variable substitution in translations

  6. Error Handling & Debugging

    • Learned to implement comprehensive error handling
    • Understood the importance of logging for production debugging
    • Discovered strategies for graceful degradation
    • Learned to handle edge cases in production environments

Project Management Learnings

  1. Iterative Development

    • Learned the value of building incrementally
    • Understood the importance of testing at each stage
    • Discovered how to pivot when technical challenges arise
    • Learned to balance feature development with stability

  2. Documentation Importance

    • Realized how critical documentation is for complex projects
    • Learned to document not just what, but why
    • Understood the value of step-by-step guides
    • Discovered how documentation helps with troubleshooting

  3. User-Centric Design

    • Learned to prioritize user needs over technical complexity
    • Understood the importance of offline functionality
    • Discovered how language barriers affect technology adoption
    • Learned to design for accessibility and simplicity

Domain-Specific Learnings

  1. Agricultural Technology

    • Learned about soil health parameters and optimal ranges
    • Understood the importance of NPK values for crop health
    • Discovered how farmers make decisions based on soil data
    • Learned about the challenges farmers face in remote areas

  2. Privacy in Agricultural Data

    • Learned about farmer concerns regarding data privacy
    • Understood the importance of on-device processing
    • Discovered how privacy-first design builds trust
    • Learned to balance functionality with privacy

  3. Accessibility in Technology

    • Learned how language barriers limit technology adoption
    • Understood the importance of offline functionality
    • Discovered how simplicity improves user adoption
    • Learned to design for users with varying technical expertise

What's next for AGROSAGE - Smart Soil Monitoring for Modern Farmers

Short-Term Enhancements (Next 3-6 Months)

  1. Enhanced AI Model

    • Train model with real field data from farmers
    • Expand training dataset with diverse soil conditions
    • Fine-tune model for better accuracy
    • Add more insight types (disease detection, pest warnings)

  2. Android ARM AI Support

    • Implement ExecuTorch for Android devices
    • Optimize for Android Neural Processing Units (NPU)
    • Test on various ARM64 Android devices
    • Achieve similar performance to iOS

  3. Additional Language Support

    • Add more regional languages (Tamil, Telugu, Marathi, etc.)
    • Expand translation coverage
    • Add language-specific agricultural terminology
    • Support right-to-left languages if needed

  4. Advanced Data Analytics

    • Historical trend analysis
    • Seasonal pattern recognition
    • Predictive insights based on historical data
    • Comparative analysis across multiple fields

  5. Crop-Specific Recommendations

    • Add crop selection feature
    • Provide crop-specific optimal ranges
    • Generate crop-specific fertilizer recommendations
    • Seasonal planting advice

Medium-Term Goals (6-12 Months)

  1. Expanded Sensor Support

    • Support for additional sensor types
    • Integration with weather APIs
    • Satellite imagery integration
    • Drone data integration

  2. Community Features

    • Farmer community platform
    • Knowledge sharing and best practices
    • Expert consultation features
    • Success stories and case studies

  3. Marketplace Integration

    • Connect farmers with fertilizer suppliers
    • Price comparison for agricultural inputs
    • Direct ordering of recommended products
    • Delivery tracking

  4. Advanced AI Capabilities

    • Computer vision for plant disease detection
    • Predictive analytics for crop yield
    • Weather-based irrigation recommendations
    • Pest and disease prediction models

  5. Enterprise Features

    • Multi-field management
    • Team collaboration tools
    • Advanced reporting and analytics
    • Integration with farm management systems

Long-Term Vision (1-2 Years)

  1. IoT Ecosystem

    • Support for multiple sensor types and brands
    • Automated irrigation systems integration
    • Smart farming equipment connectivity
    • Comprehensive farm automation

  2. Global Expansion

    • Support for more languages and regions
    • Localized agricultural knowledge
    • Regional crop database
    • Compliance with local agricultural regulations

  3. Research & Development

    • Partnership with agricultural research institutions
    • Contribution to open agricultural data
    • Development of new AI models
    • Publication of research findings

  4. Sustainability Features

    • Carbon footprint tracking
    • Water usage optimization
    • Sustainable farming recommendations
    • Environmental impact assessment

  5. Educational Platform

    • Agricultural education modules
    • Video tutorials and guides
    • Certification programs
    • Knowledge base and FAQ

Technical Roadmap

  1. Performance Optimization

    • Further reduce model size
    • Optimize inference time
    • Improve battery efficiency
    • Enhance offline capabilities

  2. Scalability Improvements

    • Support for thousands of readings
    • Efficient data compression
    • Cloud infrastructure scaling
    • Multi-tenant architecture

  3. Security Enhancements

    • End-to-end encryption
    • Biometric authentication
    • Secure data backup
    • Privacy-preserving analytics

  4. Platform Expansion

    • Web application version
    • Desktop application
    • Smartwatch integration
    • Voice assistant integration

Community & Impact Goals

  1. Farmer Empowerment

    • Reach 100,000+ farmers in first year
    • Improve crop yields by 20% on average
    • Reduce fertilizer costs by 15%
    • Save water through optimized irrigation

  2. Open Source Contribution

    • Open source core components
    • Contribute to agricultural AI research
    • Share best practices and learnings
    • Build developer community

  3. Partnerships

    • Collaborate with agricultural universities
    • Partner with government agricultural departments
    • Work with NGOs for rural outreach
    • Integrate with existing farm management systems

  4. Sustainability Impact

    • Reduce water waste through optimized irrigation
    • Minimize fertilizer overuse
    • Promote sustainable farming practices
    • Contribute to food security

Conclusion

AgroSage represents a significant step forward in making agricultural technology accessible, affordable, and practical for farmers worldwide. By combining cutting-edge ARM-based AI with offline-first design and multi-language support, we've created a tool that truly serves farmers in remote areas with limited resources.

The project demonstrates that with the right technology choices and user-centric design, we can build solutions that bridge the digital divide in agriculture, empowering farmers to make data-driven decisions and improve their livelihoods.

Our mission continues: To democratize agricultural technology and make smart farming accessible to every farmer, regardless of their location, language, or technical expertise.

Built With

  • bolt.new
  • cursor
  • dht11
  • esp32
  • executorch
  • expo.io
  • npk
Share this project:

Updates