Pungde

Empowering the hands that feed the world.

Comment

Pungda - AI-Powered Farming Assistant 🌾

Inspiration

Millions of farmers worldwide struggle with basic questions: "Can I grow this crop here?" "Which seeds should I buy?" "How can I increase my yield?" We built Pungda (meaning "field" in Garhwali) to democratize agricultural knowledge using Google Cloud's AI and satellite technology, making expert farming guidance accessible to every farmer regardless of location or resources.

What it does

Pungda is an AI-powered farming assistant providing intelligent crop cultivation decisions through conversational AI:

Core Features:

  • Crop Yield Prediction: ML model trained on 50,000+ global data points with real-time Google Earth Engine satellite imagery (64-dimensional embeddings), predicting yields for 10 major crops (rice, wheat, maize, cotton, coffee, banana, coconut, chickpea, kidneybeans, lentil, pigeonpeas)
  • Multi-Agent AI System: Six specialized agents (Agri Analyzer, Crop Suitability, Grow Anyways, Yield Improvement, Seed Identifier, Image Generator) working seamlessly together
  • Real-Time Intelligence: Google Alpha Earth Dataset (yearly embeddings from multiple satellites), NASA POWER API, Google Geocoding, and 12-month climate analysis
  • Farmer-Friendly Interface: Next.js web app with three themes, real-time chat, PDF export, and mobile-responsive design

Farmers ask natural questions and receive comprehensive, actionable answers with yield predictions, climate analysis, specific seed recommendations with buying links, improvement strategies, and visual guides.

How we built it

Architecture: Distributed microservices on Google Cloud Run

Tech Stack:

  • ML Pipeline: XGBoost regression, Google Alpha Earth Dataset (64-dim yearly embeddings), NASA POWER API, Scikit-learn, SPAM 2020 + Kaggle crop datasets
  • Backend: FastAPI (ML service), Google ADK (multi-agent orchestration), Gemini 2.5 Flash (6 AI agents), Vertex AI Imagen 4.0 (image generation)
  • Frontend: Next.js 16, TypeScript, Tailwind CSS, React Markdown, jsPDF
  • Infrastructure: Cloud Run (3 microservices with auto-scaling), Cloud Storage, Vertex AI, Earth Engine, Geocoding API

Development Process:

  1. Data & Training: Combined SPAM 2020, Kaggle datasets, and Google Alpha Earth embeddings; trained XGBoost on 50,000+ samples
  2. ML Service: FastAPI service with geocoding, Earth Engine integration, NASA POWER API, and XGBoost inference
  3. Multi-Agent System: Built hierarchical agent system with ADK - root orchestrator + 6 specialized sub-agents with Vertex AI Imagen integration
  4. Web Interface: Next.js app with SSR, streaming responses, localStorage sessions, 3 themes, and PDF export
  5. Deployment: All services on Cloud Run with proper authentication, CORS, and auto-scaling

AI Studio Partnership: Google AI Studio was transformative - I focused on architecture, datasets, and design decisions while AI Studio generated production-ready code, FastAPI endpoints, React components, and deployment configs. True "vibe coding" - turning ideas into reality through AI collaboration.

Challenges

  1. Google Alpha Earth Dataset: Brand new technology with sparse documentation and few examples - spent days understanding data structure and efficient querying
  2. Agricultural Data: Evaluated dozens of fragmented datasets before successfully combining SPAM 2020, Kaggle requirements, and Alpha Earth embeddings
  3. CORS with ADK: ADK's automatic Cloud Run deployment gave no CORS control - solved by configuring at service level
  4. PDF Generation: Converting markdown with images to professional PDFs while handling async loading and mobile constraints
  5. Outdated ADK Documentation: Many tutorials deprecated - relied on open-source GitHub examples for current best practices
  6. Multi-Agent Coordination: Implemented strict orchestration pattern for reliable agent collaboration
  7. Image Placeholder Conversion: Refined root agent prompts with explicit instructions and regex for reliable placeholder handling
  8. Response Time: Optimized from 15-20s to 8-12s through streaming, parallel generation, and prompt optimization

Accomplishments

  • Sophisticated Multi-Agent System: Six specialized agents collaborating seamlessly using Google ADK
  • Pioneering Alpha Earth Integration: Among first to use Google Alpha Earth Dataset's satellite-trained yearly embeddings in production ML
  • Global Scalability: Works for ANY location worldwide with automatic local adaptation
  • Farmer-Friendly UX: Simple conversations despite complex technology, with 3 beautiful themes
  • Actionable Guidance: Real buying links, specific varieties, cost estimates, implementation steps
  • Visual Learning: Contextual images for better understanding, especially for low-literacy users
  • Production-Ready: Cloud Run deployment with auto-scaling, error handling, monitoring, security
  • Comprehensive Documentation: Detailed READMEs, architecture diagrams, code documentation
  • ML Accuracy: Reliable predictions from 50,000+ global locations with satellite embeddings
  • Full Google Cloud Integration: Cloud Run, Vertex AI (Gemini + Imagen), Earth Engine, Geocoding, Storage, ADK

What we learned

Technical:

  • Google ADK is powerful for multi-agent systems with proper prompt engineering
  • Cloud Run perfect for AI workloads with auto-scaling and service authentication
  • Google Alpha Earth Dataset revolutionary but requires patience with limited documentation
  • Gemini 2.5 Flash provides excellent speed/intelligence balance with streaming
  • Vertex AI Imagen creates impressive agricultural images with proper prompting

Domain:

  • Agriculture is complex and context-dependent - personalization is crucial
  • Farmers need actionable guidance with specific steps, not just data
  • Visual communication critical for engagement and understanding
  • Trust requires transparency - show data and explain reasoning

Development:

  • AI Studio transforms development - focus on thinking/designing while AI handles implementation
  • Microservices enable parallel development and independent scaling
  • Documentation essential for complex systems - document while building
  • Graceful error handling makes or breaks UX
  • Global testing reveals important edge cases

What's next

Immediate (3 Months):

  • Revolutionary ML Model: Cluster-based approach using Alpha Earth embeddings to identify similar locations globally - predict ANY crop/vegetable by learning from similar environments worldwide (targeting 90%+ accuracy)
  • Voice Calling Agent: Real-time multilingual voice interface for older, less tech-literate farmers - works on any phone including feature phones
  • Enhanced Data: Expand NASA POWER usage, integrate FAO soil maps, market prices, weather forecasts, pest databases
  • GPU Inference: Cloud Run with NVIDIA L4 GPUs for 10x faster predictions and real-time voice processing

Medium-Term (6-12 Months):

  • Native mobile apps with offline capabilities, camera integration, voice interface, SMS/WhatsApp
  • Community features: forums, success stories, expert network, marketplace, cooperative tools
  • Analytics dashboard: yield tracking, comparative analysis, ROI calculations, climate projections
  • Precision agriculture: IoT sensors, drone imagery, automated irrigation, variable rate fertilization

Long-Term (1-2 Years):

  • AI Farming Ecosystem: Supply chain integration, financial services, equipment rental, training, government integration
  • Climate Resilience: Adaptation strategies, drought/flood-resistant crops, carbon tracking, sustainability incentives
  • Global Expansion: 50+ languages, partnerships with agricultural ministries, offline-first architecture
  • Research Collaboration: University partnerships, open-source components, UN FAO initiatives

Impact Goals by 2027:

  • Serve 10M farmers across 50+ countries
  • Improve yields by 20-30% through optimized practices
  • Reduce crop failures by 40% through better suitability analysis
  • Save farmers $500M annually through better decisions
  • Contribute to global food security and reduce environmental impact

Agriculture feeds the world, yet farmers lack access to transformative knowledge. Pungda democratizes agricultural expertise through AI, making world-class farming guidance accessible to anyone with a smartphone. With Google Cloud's infrastructure and AI, we're building a movement toward intelligent, sustainable, and equitable agriculture for the 21st century. 🌾🚀

Detailed overview: https://github.com/luexclothings-hue/Pungde/blob/main/HACKATHON_SUBMISSION.md

Built with ❤️ for farmers worldwide using Google Cloud Run, Vertex AI, Earth Engine, and ADK.

Built With

Share this project:

Updates