FarmAgent

• 

  Dashboard

• 

  Architecture

Inspiration

Farm advice tools are opaque. We wanted an inspectable agent that shows how it reasons, not just the answer—so farmers/researchers can trust and verify decisions.

What it does

Takes a crop question + optional image, builds a plan, runs tools (crop ID, disease hint, weather/soil), then uses Gemini to produce a concise, evidence-backed recommendation. Live dashboard shows plan & final output.

How we built it

ADK Sequential graph: Planner → Governor → Executor → Synthesizer. Planner creates JSON plan. Governor (callbacks) validates/guards. Executor calls Python tools; receipts logged server-side. Synthesizer (Gemini) writes the final answer. Two Cloud Run services: Flask frontend and ADK backend. Vertex AI for model access.

Challenges we ran into

Agent loops/over-reasoning, and balancing UI clarity with transparency. Service-to-service config on Cloud Run (env/secrets, timeouts) and stabilizing logs under load.

Accomplishments that we're proud of

End-to-end explainable agent on Cloud Run with clean ADK wiring, fast cold-start, and a simple gateway that keeps the system stateless and scalable.

What we learned

ADK makes agent graphs testable; callbacks (Governor) are the right place for safety nudges. Serverless + Vertex AI is a fast path from prototype to demo.

What's next for FarmAgent

We plan to integrate real farm sensor data, expand multilingual support for regional farmers, and add offline recommendations powered by cached Gemini insights.

Built With

• google-adk
• javascript
• python