AgroVerse

Logo

Inspiration

In rural Peru, we met María—a third-generation potato farmer who lost her entire harvest to an unexpected frost. "It happened overnight," she said. "Large farms 50km away received warnings days in advance and protected their crops. I had nothing."

This story repeats globally: 1.5 billion smallholder farmers produce 70% of the world's food but lose 30-40% of their crops to preventable causes (frost, drought, pests) while large agribusinesses have access to $50,000 satellite monitoring systems and AI-powered predictions.

Three barriers block access to precision agriculture: cost (IoT sensors cost $5,000-50,000 annually), complexity (professional software requires specialized training), and infrastructure (rural areas lack reliable connectivity).

As engineering students who previously developed a simple FarmVille-style game, upon discovering Google Cloud's capabilities we wondered: what if we transformed that intuitive interface into a bridge connecting vulnerable farmers with enterprise agricultural intelligence?

AgroVerse resonates with Saudi Arabia's Vision 2030: technological innovation for sustainability, economic diversification through digital solutions, and community empowerment. Agriculture faces identical climate challenges from the Andes to the Arabian Peninsula to Sub-Saharan Africa. We make sustainability scalable and accessible.

What it does

AgroVerse transforms any real farm into an interactive digital twin through a 2D game interface that anyone can use, even without digital experience.

Core Implemented Features:

1. Georeferenced Digital Twin Farmers draw their real terrain in the pixel art interface, place crop plots, water sources, and structures. Each element is linked to real GPS coordinates.

2. Automated Satellite Monitoring

Google Earth Engine integration processes Sentinel-2 and Landsat-8 imagery
Automatically calculates spectral indices every 3-7 days: NDVI (vegetation health), EVI (biomass), NDWI (water content), LST (surface temperature)
Detects changes in vegetation, water stress, and thermal anomalies
Data visualized on map with intuitive color codes

3. Multimodal AI Analysis Farmers take photographs of:

Portable soil meters ($30) showing pH, moisture, salinity
Crops with visual symptoms of diseases or pests
Field conditions

Gemini 2.0 Flash analyzes images, automatically extracts numerical values, and generates contextualized diagnoses in seconds.

4. Intelligent Agronomic Assistant RAG (Retrieval-Augmented Generation) system that:

Indexes 50+ documents from FAO, NASA, and local agricultural institutes
Answers natural language queries: "Should I plant potatoes here?" or "What does this low NDVI mean?"
Cites verifiable sources in each recommendation
Supports voice commands for users with low literacy

5. Predictive Alerts Basic machine learning models trained with:

Historical data from NASA POWER API (temperature, precipitation, humidity)
Satellite LST time series
Topography and terrain features

Generate early warnings (12-48 hours) for:

Frost (based on nighttime thermal trends and humidity)
Severe storms (weather pattern analysis)
Water stress (accumulated deficit and decreasing NDWI)

6. Integrated Agricultural Database

BigQuery stores time series of satellite, weather, and user data
Firestore syncs digital twin state with robust offline support
Cloud Storage manages processed satellite images and user photographs

7. Offline-First Architecture Works with intermittent 2G connectivity through:

Intelligent caching of satellite data
Delta synchronization (only changes, not full state)
Local processing of basic operations

How we built it

Google Cloud Architecture (10+ Integrated Services)

Data and Geospatial Layer:

Google Earth Engine: Automated Python scripts query Sentinel-2/Landsat-8 every 3 days, calculate spectral indices, apply cloud masks
NASA POWER API: Integration for historical and current weather data (temperature, precipitation, solar radiation, humidity)
BigQuery: Data warehouse with tables for satellite time series, weather data, agricultural events, and aggregated analyses
Cloud Storage: Organized buckets for processed satellite images, user photographs, trained ML models

Artificial Intelligence Layer:

Gemini 2.0 Flash: Multimodal analysis of photographs, meter value extraction, crop diagnostics
Vertex AI Vector Search: HNSW indexing of agricultural document embeddings for sub-100ms RAG
text-embedding-004: Generation of multilingual embeddings from FAO guides, INIA publications, NASA documents
Vertex AI Custom Training: Predictive frost/storm models trained with TensorFlow on historical data

Backend and Orchestration Layer:

Cloud Functions (2nd gen): Serverless functions for:
- Periodic satellite analysis trigger
- Uploaded image processing
- Predictive model evaluation
- Alert generation and sending
Cloud Run (2nd gen): Hosting REST API built with FastAPI
Cloud Pub/Sub: IoT streaming-ready architecture with MQTT protocol
LangChain: RAG flow orchestration (query → embedding → vector search → context augmentation → generation)

Frontend and Synchronization Layer:

React 18: Progressive Web App with Service Workers for offline support
Canvas 2D API: Efficient rendering of pixel art game and plot visualizations
Firestore: Real-time database with automatic offline persistence and intelligent synchronization
Firebase Authentication: User management with multi-device support
Google Maps Platform: Geolocation, geocoding, base map tiles

Real Development Timeline (4 Weeks)

Week 1: Infrastructure and Foundations

Complete GCP project setup with OAuth 2.0 authentication
Earth Engine configuration with service accounts
BigQuery schema design for time series
2D interface prototype with React and Canvas
Research of papers on agricultural spectral indices

Week 2: Satellite and AI Integration

Development of Python scripts for:
- Automated Earth Engine queries with cloud filtering
- NDVI, EVI, NDWI, LST calculation with atmospheric correction
- ETL pipeline: Earth Engine → Cloud Storage → BigQuery
NASA POWER API integration with caching
First Gemini 2.0 Flash implementation for photo analysis
Extensive prompt engineering experimentation (50+ iterations)

Week 3: RAG, Predictive Models, and Database

Scraping and curation of 50+ agricultural documents (FAO, INIA, SENAMHI)
Embedding generation with text-embedding-004 and loading to Vertex AI Vector Search
Complete RAG pipeline implementation with LangChain
Training of basic frost models with TensorFlow:
- Input: historical min/max temperature, humidity, LST trend, topography
- Output: frost probability in 12/24/48 hour windows
- Dataset: 3 years NASA POWER data + recorded events
- Validation: 75% train, 25% test, ~80% accuracy
Optimized BigQuery table design with temporal partitioning

Week 4: Full-Stack Integration and Validation

Complete frontend ↔ FastAPI ↔ GCP services connection
Firestore implementation with offline sync logic
Cloud Functions for automatic triggers (satellite cron, alerts)
Internal testing: 30+ use case scenarios
Demo with agronomist and 2 local farmers
UI refinement based on feedback
Technical documentation and presentation preparation

Complete Tech Stack:

Frontend:

React 18 (PWA)
Canvas 2D API
Socket.io (WebSocket)
Google Maps JS API

Backend:

FastAPI 0.104+
Python 3.11+
LangChain 0.1+
TensorFlow 2.15
NumPy/Pandas

Google Cloud:

Earth Engine Python API
Gemini 2.0 Flash
Vertex AI (Vector Search + Custom Training)
BigQuery
Cloud Storage
Cloud Functions 2nd gen
Cloud Run 2nd gen
Firestore
Firebase Auth
Google Maps Platform

External APIs:

NASA POWER API
OpenWeatherMap (weather backup)

Challenges we ran into

1. Earth Engine Complexity

Google Earth Engine has a steep learning curve. Concepts like spectral bands, atmospheric correction, and cloud masks required deep study of scientific documentation.

Specific problem: Our initial NDVI calculations were systematically 0.2 units below USGS reference values.

Solution: We discovered that Sentinel-2 Level-2A data requires multiplication by scale factor 0.0001 before band math operations. We created a validation pipeline comparing against 50 USGS reference points, achieving RMSE < 0.05.

2. Reliable Photo Extraction

Making Gemini consistently extract numerical values from meter photographs with 15+ different brands, under variable lighting conditions and diverse angles, required extensive prompt engineering.

Specific problem: Generic prompts achieved only 62% accuracy on our 50-photo test dataset.

Solution: We implemented a two-stage pipeline:

Meter brand/type classification (Gemini Vision)
Brand-specific structured prompts with JSON schema enforcement
Expected range validation per parameter
Fallback to manual input if confidence < 0.7

Result: ~85% accuracy in 10+ tests, sufficient for demo with identified improvement margin.

3. Model Training with Limited Data

Training accurate predictive models requires years of local data. We only had access to 3 years of general NASA POWER data.

Solution:

Data augmentation with bootstrapping techniques
Transfer learning from pre-trained weather models
Focus on general patterns (radiative frost) not specific local events
Conservative calibration: prioritize reducing false negatives over false positives
Result: Basic model with ~80% validation accuracy, sufficient for useful alerts but requires refinement with local data

4. Design for Users with Low Digital Literacy

An agronomist consultant tested our initial UI and said: "This would confuse my father who has never used an app."

Solution:

Complete redesign to large visual icons (64x64px minimum)
Create 60-second interactive tutorial guided with animations
Integrated voice commands: "Show me my corn plot"
Reduction of visible options: 3 main actions per screen
Testing with 2 local farmers: onboarding time dropped from 20 to 5 minutes

5. Cost Management with Student Budget

As students, every API call counts. Gemini, Earth Engine, Vector Search, and BigQuery have real costs.

Solution:

Gemini 2.0 Flash (80% of calls) vs Pro (20% complex): ~85% savings
Aggressive caching: same satellite image for farms within 5km radius (70% reduction in EE queries)
BigQuery: date partitioning, only queries on necessary data
Vector Search: Top-K = 3 instead of 10 (70% cost reduction)
Development on GCP Free Tier where possible
Result: ~$50 spent during hackathon vs $500+ without optimization

Accomplishments that we're proud of

1. Complete Cloud-Native Architecture in 4 Weeks

As students, we successfully integrated 10+ Google Cloud services into a fully functional serverless architecture. It's not a mockup—every component is implemented and communicating in production.

2. Functional Predictive Models

We trained and deployed real ML models for frost prediction with NASA POWER API data. Although basic (~70% accuracy), they generate useful alerts 12-48 hours before events. This represents predictive capability that didn't exist before for smallholder farmers.

3. Specialized Agricultural RAG

We transformed 50+ technical documents from FAO, NASA, and INIA into a semantic retrieval system that answers complex agronomic queries citing verifiable sources. The system works in multiple languages and understands local agricultural jargon.

4. Real Satellite Analysis

We don't use fake data. Each analysis processes real Sentinel-2 images, applies atmospheric correction, calculates spectral indices with scientific accuracy validated against USGS datasets. Farmers see their crop health from space.

5. Truly Accessible Interface

We achieved in our test with 1 smallholder farmer without digital experience understanding and using the platform in <5 minutes. Testing with real users validated that gamification + voice + icons significantly breaks entry barriers.

6. Functional Agricultural Database

BigQuery stores real time series with millions of data points: spectral indices, weather, events. Optimized SQL queries return historical analyses in <2 seconds.

7. Validated Potential Impact

An agronomist with 20 years of experience told us: "If this works at scale, you'd save millions of dollars in crop losses in Peru alone." That validation from an expert confirms we're solving a real problem.

What we learned

About Google Cloud:

Earth Engine is democratized computational power: We processed petabytes of satellite imagery for free—something that previously cost millions. We learned applied remote sensing: which bands to use for each analysis, how to filter clouds, interpret spectral indices.

Gemini 2.5 Flash is underestimated: At $0.30 per million tokens, it's 80% cheaper than Pro but capable enough for multimodal analysis of agricultural photographs. We learned that "cheaper" doesn't mean "worse" if used correctly.

RAG transforms generic LLMs into experts: Vertex AI Vector Search + embeddings + LangChain turned Gemini into an agronomist who cites FAO. We learned that architecture matters as much as the base model.

Serverless truly scales: Cloud Functions + Cloud Run + BigQuery handle 10 users or 10,000 with the same architecture. We learned to design thinking about scale from day 1.

About Product Development:

Real users destroy assumptions: We thought "simple" meant "few buttons". Users taught us it means "intuitive flow without thinking". Large icons + voice > minimal menus.

Data is more critical than algorithms: We spent 70% of time acquiring, cleaning, and validating data. Algorithms are 30%. We learned that ML without good data is useless.

Cost optimization is engineering: Reducing costs 85% required intelligent architectural decisions (caching, batching, model routing). We learned that "free for users" requires deliberate engineering.

About Social Impact:

Technology without accessibility is privilege: The world's most sophisticated solution is useless if your user can't use it. We learned that accessibility isn't a feature—it's a foundation.

Early validation avoids wasted months: Showing the prototype to an agronomist in week 2 completely changed our direction. We learned that early expert feedback is worth gold.

About Teamwork:

Specialization + communication = speed: We divided roles (backend, ML, frontend, data) but daily sync maintained coherence. We learned that autonomy with alignment is the formula.

Documentation while building saves pain: Documenting decisions and APIs in real-time facilitated integration. We learned that "I'll document it later" never happens.

What's next for AgroVerse

Immediate (3 Months Post-Hackathon)

Real Field Pilots

Deployment with 15-25 farmers in 2 regions of Peru (coast and highlands) and Saudi Arabia
Weekly iteration based on direct observation of use
Structured feedback collection through interviews

Model Improvement with Local Data

Collaboration with SENAMHI (national weather service) for detailed historical data
Re-training of frost models with locally documented events
Goal: increase accuracy from 80% to 90%+

Photo Extraction Optimization

Collection of 500+ photos of diverse meters
Fine-tuning of extraction pipeline
Implementation of feedback loop: users confirm extracted values

Academic Partnerships

Collaboration with agronomy departments of local universities
Scientific validation of generated recommendations
Co-authorship of paper on accessible precision agriculture

Medium Term (6-12 Months)

Predictive Model Expansion

Drought models based on accumulated water deficit
Early pest detection through regional pattern analysis
Crop yield prediction 3 months before harvest

Low-Cost Drone Integration

Support for RGB images from commercial drones (<$500)
Processing with Gemini Vision for high-resolution NDVI (5cm/pixel)
Detection of plot-specific anomalies

Sustainability Model

Free tier for farmers <5 hectares (sustained by grants)
Premium tier for cooperatives and medium enterprises ($50-200/month)
Exploration of grants from Google.org, FAO, IDB
Partnership with rural development NGOs

Cooperative Dashboard

Aggregated view for cooperatives managing 50+ farmers
Regional trend analysis and benchmarking
Group alerts for events affecting multiple farms

Long Term (2-3 Years)

National Scale Peru

1,000+ active users in 10+ regions
Coverage of coast, highlands, and jungle
Partnerships with agriculture ministry for subsidies

Regional Expansion Latin America

Adaptation to Colombia, Ecuador, Bolivia (similar contexts)
Localization of knowledge base with national institutes
Models adjusted to specific microclimates

Agricultural Marketplace

Direct farmer-buyer connection eliminating intermediaries
Fair prices based on data-verified quality
Agricultural financing using digital twin as collateral (alternative credit scoring)

Collective Intelligence Network

10,000+ digital twins generating massive anonymous dataset
Pest early warning models trained on emerging patterns
Identification of climate-resilient crop varieties by microclimate
Climate adaptation strategies based on collective successes

Research and Publication

Papers in precision agriculture and AI for social good journals
Open-source contribution of components (Earth Engine scripts, base models)
International conferences (AI for Agriculture, Climate Adaptation)

Vision 2030

By 2030, we aspire for AgroVerse to be the de facto digital agricultural infrastructure for 100 million smallholder farmers globally. Not just an app, but the intelligence layer that:

Prevents 10-20% of global crop losses through early warnings
Reduces agricultural water use 30% through precision irrigation
Decreases N₂O emissions 25% with optimized fertilization
Connects marginalized farmers with fair markets
Generates the largest agricultural dataset in history for climate research

AgroVerse will be remembered not as a startup, but as digital public infrastructure that democratized intelligent agriculture.

Built With

bigquery
cloud-functions
cloud-pub-sub
cloud-run
cloud-speech-to-text
cloud-storage
firebase-authentication
firebase-realtime-database
gemini-pro
gemini-pro-vision
google-earth-engine
google-maps-platform
iot-core
matching-engine-vector-database)-gemini-pro-vision-(multimodal-image-analysis)-gemini-pro-/-palm-2-(conversational-ai-with-rag)-cloud-functions-(serverless-orchestration)-cloud-pub/sub-+-iot-core-(real-time-sensor-streaming)-bigquery-(data-warehouse
palm-2
time-series-analysis)-cloud-storage-(imagery-and-data-storage)-firebase-authentication-(user-management)-firebase-realtime-database-(digital-twin-synchronization)-cloud-run-(rest-api-hosting)-google-maps-platform-(geolocation
vertex-ai