Shamba Steward

One messy field note in; a Gemini agent on Cloud Run parses, plans, VERIFIES against agrochemical safety intervals, and delivers a calendar + market message.

Comment

The problem

A smallholder farmer holds a dozen half-finished decisions at once - what to spray, what is ripe, what the weather will do, what to sell. Nothing sequences them, and nothing catches the unsafe move: spraying a crop you are about to harvest, inside its chemical pre-harvest interval.

What Shamba Steward does (Taskmaster track)

Type ONE messy field note and the agent runs a real, autonomous workflow - it takes action, it does not just chat:

  1. Extract the note into structured field-events (Gemini 2.5 Flash on Vertex AI).
  2. Recall the farm's prior context from the Firestore Memory Bank.
  3. Plan - schedule each action; treatments precede harvest; never spray on a forecast rain day.
  4. Verify - the honest safety layer: it flags any spray inside its agrochemical pre-harvest interval before a same-crop harvest, and refuses to present an unsafe plan.
  5. Deliver - a downloadable calendar (.ics) of the work and a ready-to-send market message.

Architecture (Google Cloud, decoupled)

  • Google ADK LlmAgent on Cloud Run (scales to zero).
  • Gemini 2.5 Flash via Vertex AI - reasons and routes; the key never leaves the project (ADC, no API key in code).
  • Deterministic Python tools do the load-bearing work (scheduling + the pre-harvest safety check), unit-tested independently - an LLM cannot hallucinate a safe plan past the verifier.
  • Firestore Memory Bank holds cross-session farm context; memory failures degrade gracefully.

Why it wins the judging

  • Innovation & Operational Utility: it acts - a verified schedule + artifacts from one note, catching a real agronomic error a chatbot would miss.
  • Architectural Discipline: Gemini reasons, deterministic tools compute, Firestore persists; credentials via Vertex ADC; graceful degrade.
  • Demo & Production Readiness: reproducible README, an architecture diagram, and a live agent on Cloud Run + Vertex AI.

How we built it

Google ADK + Gemini (Vertex) + Cloud Run + Firestore. Deterministic safety core in pure Python, unit-tested green. Built and hardened with amini-code multi-agent runs. Code and spin-up instructions in the repo.

Built With

  • cloud-run
  • firestore
  • google-adk
  • google-gemini
  • python
  • vertex-ai
Share this project:

Updates