CropPilot

Inspiration

Agriculture is one of the most data-rich industries in the world, yet most small farmers still make decisions based on intuition and experience alone. We wanted to change that. Seeing how unpredictable weather, soil degradation, and volatile crop prices affect farmer livelihoods inspired us to build something that could put real-time, actionable intelligence directly in a farmer's hands, without requiring any technical expertise.

What it does

CropPilot is an AI-powered farming co-pilot that helps farmers make smarter decisions in real time. You can ask it anything, whether that is whether to irrigate today, what is wrong with a sick plant, or when the best time to sell your harvest is. It pulls live weather forecasts, analyzes soil health data, calculates precise irrigation needs, and tracks commodity market prices. You can also upload a photo of a diseased crop and the AI will diagnose it instantly. Everything is surfaced through a clean chat interface with a live farm dashboard that updates after every query.

How we built it

We built CropPilot as a full-stack Next.js web app deployed on Vercel. The core of the app is an agentic AI system powered by Groq and Llama 3.3 70B, which uses tool calling to pull data from multiple sources before forming a response. We integrated the Open-Meteo API for real weather and forecast data, built a soil sensor simulation model, and added a commodity price engine for market recommendations. For crop disease detection we used Llama 4 Scout, a vision model, which lets the agent analyze uploaded plant photos. Voice input was added using the browser's built-in Web Speech API.

Challenges we ran into

The biggest challenge was getting the agentic tool-calling pipeline to work reliably. When the model runs low on tokens it sometimes generates malformed function calls, which took a while to debug and resolve by switching models and tuning the system prompt. We also ran into a deprecated vision model mid-build and had to quickly find and switch to a supported replacement. Getting the dashboard to sync with the chat state in real time also required some careful state management in React.

Accomplishments that we're proud of

We are proud that CropPilot actually works end to end. The agent reasons across multiple live data sources in a single response, the vision model correctly identifies plant diseases from photos, and the dashboard updates dynamically with every query. Building a fully deployed, production-ready app in under 24 hours with this many integrated systems feels like a real accomplishment.

What we learned

We learned a lot about building reliable agentic pipelines, specifically how much prompt design and model selection matters when chaining tool calls. We also got hands-on experience with multimodal AI, vision models, and how to structure a Next.js app for both frontend and backend in one codebase. Most importantly we learned how to ship fast without sacrificing quality.

What's next for CropPilot

The next step is connecting CropPilot to real IoT soil sensors so the data is live rather than simulated. We also want to add multilingual support so farmers in India, Africa, and Southeast Asia can use it in their native language. Longer term we see CropPilot becoming a full farm management platform with historical tracking, yield prediction, and integration with irrigation hardware for fully autonomous control.

Built With

• css
• css-in-js
• github
• groq
• javascript
• llama-3.3-70b
• llama-4-scout-17b-vision
• next.js
• open-meteo-forecast
• tailwind
• vercel
• web-speech-api