AgroLatam Agent

Inspiration 70% of Latin American farmers make critical decisions — when to harvest, when to sell, when to apply for credit — without access to real-time market data. Large corporations have sophisticated data systems. Small farmers have nothing but rumors and outdated prices. We saw an opportunity to level the playing field. With 60 million smallholder farmers generating over $200 billion in annual exports across 18 countries, the potential impact was massive. AgroLatam Agent was built to give every farmer — regardless of size or location — the same intelligence that moves global commodity markets.

What it does AgroLatam Agent is an autonomous AI platform that monitors, reasons and acts on behalf of Latin American farmers without human intervention.

Monitors real-time prices for 11 crops (coffee, cacao, corn, avocado, soy, palm oil, rice, sugarcane, banana, orange and tomato) across ICE NY, ICE London, CME, FAO and BMD exchanges Forecasts agricultural weather with 7-day forecasts, soil conditions and harvest risk alerts for 18 countries Acts autonomously — sending price alerts, generating credit documents, identifying export buyers and recommending optimal harvest and sell windows Calculates farm profitability using real market prices with crop comparison Shows local markets, cooperatives and exporters near the farmer's location on an interactive map Provides an intelligent agricultural calendar adapted by crop and country

How we built it AI Agent: Gemini AI + Groq/Llama 3.3 for ultra-fast autonomous reasoning and multi-step planning Data Pipelines: Fivetran MCP synchronizes commodity prices, weather data and export statistics every 5 minutes from global sources Backend: FastAPI + Python deployed on Hugging Face Spaces, exposing REST endpoints for prices, weather, alerts and chat Database: Supabase (PostgreSQL) with row-level security storing farmer profiles, alerts and historical price data Weather: Open-Meteo API providing free, real-time agricultural forecasts for all LATAM regions Frontend: 14 HTML/CSS/JS pages deployed on Vercel including dashboard, analytics, map, calendar, calculator, news, weather and admin panel — fully bilingual (EN/ES) and mobile responsive

Challenges we ran into Gemini API quota: The free tier quota was quickly exhausted under load. We migrated to Groq + Llama 3.3 which proved 10x faster and more reliable for our use case. Dependency conflicts: Supabase and httpx had incompatible version requirements that required careful resolution across different environments (local vs Hugging Face Docker). Heterogeneous LATAM data: Each of the 18 countries has different data formats, government APIs and agricultural reporting standards. Normalizing this through Fivetran required significant schema mapping work. Hugging Face cold starts: Free tier spaces sleep after inactivity, causing 30-60 second delays. We implemented a wake-up ping on page load and added a graceful offline state in the frontend. Mobile responsive navbar: With 7+ tool pages, fitting the navigation cleanly on mobile required redesigning the entire navbar with a hamburger menu and organized tool grid.

Accomplishments that we're proud of

Built a truly autonomous agent that acts without being asked — not just a chatbot Covered 18 countries and 11 crops in a single unified platform Achieved full bilingual support (EN/ES) with dynamic language switching across all 14 pages Deployed a complete production stack entirely on free services (Vercel + Hugging Face + Supabase) Built an interactive map that finds real markets, cooperatives and exporters near the farmer using OpenStreetMap and the Overpass API Created an agricultural calendar with crop-specific planting, care and harvest schedules adapted by country

What we learned

How to design truly autonomous multi-step agents that go beyond question-answering into real-world action The power of Fivetran MCP for building reliable data pipelines that connect heterogeneous agricultural data sources That the most impactful AI applications are not always the most glamorous — sometimes the biggest opportunity is in the markets that technology has consistently overlooked How to build and deploy a full-stack AI product entirely on free infrastructure without compromising on features or reliability That language matters — building in Spanish from the start dramatically increases accessibility for the target audience

What's next for AgroLatam Agent 📡 Real-time commodity prices Connect live price feeds directly from ICE NY, ICE London, CME and FAO exchanges to replace the current static data with true real-time pricing updated every 5 minutes. 📱 WhatsApp Business API Send autonomous price alerts, weather warnings and harvest recommendations directly to farmers' phones via WhatsApp — the most used communication app in Latin America with over 85% penetration. 📄 PDF Document Generator Automatically generate agricultural credit applications, production reports and export letters that farmers can download and present directly to banks and cooperatives — eliminating bureaucratic barriers to credit access. 🔔 Browser Push Notifications Alert farmers instantly on any device without requiring them to open the platform — critical for time-sensitive market opportunities. 🤝 Cooperative Management Allow farming cooperatives to manage members, aggregate selling power and negotiate better prices collectively with exporters. 📴 Offline PWA Mode Progressive Web App support so farmers in remote areas with limited or no connectivity can still access key market data and recommendations. 🌍 Expansion beyond LATAM The same information gap exists for smallholder farmers across Sub-Saharan Africa and Southeast Asia. AgroLatam Agent's architecture is designed to scale globally.

Built With

• bigquery
• cloud-run
• cloud-storage
• fao-api
• fivetran
• fivetran-mcp
• gemini
• google-cloud
• google-cloud-agent-builder
• html
• ice-futures-api
• javascript
• model-context-protocol
• open-meteo-api
• python
• rest-api
• secret-manager
• whatsapp-business-api