Inspiration

Smallholder farmers make daily decisions — water, fertilizer, pests — with no agronomist on call and often no signal in the field. Harvest Notes puts a patient, private advisor in their pocket, works where the network does not, and never sends a note to the cloud unless the farmer's request needs it.

What it does

Harvest Notes is an offline-first AI field journal for smallholder farmers. A farmer writes a short field note — crop, what they saw, what worries them — and the app turns it into useful work: a plain-language summary, concrete tasks to do, answers drawn from their own notes, and a 4-week season plan.

How we built it

Hybrid AI is the core of the app. It uses Firebase AI Logic with one hybrid model: short, private notes go to Gemini Nano on-device when the model is available in Chrome; longer prompts and season plans go to Gemini Flash in the cloud; and when the device is offline, on-device inference keeps working with no network at all. Every AI answer shows an honest source badge ("Gemini Nano · on-device", "Gemini Flash · cloud", or "DEMO answer") so the farmer always knows where the answer came from. If the cloud is overloaded, the app retries the fallback model; if a call fails, it says so instead of inventing an answer.

The app is offline-first end to end: Firestore with offline persistence, so notes are written in the field with no signal and sync when the farmer is back in range. Anonymous Authentication keeps each farmer's notes private (Firestore rules: owner only). Remote Config controls the cloud model and limits without a redeploy. App Check is wired in. It is a zero-build static PWA that installs from the browser and works on a cheap Android phone — the device most smallholder farmers actually carry.

Challenges we ran into

The cloud model we first picked was deprecated mid-build and started returning errors, so we re-routed to a supported model and made the cloud model a Remote Config value. A service-worker cache also kept serving the old app shell after deploys, so we switched to a versioned cache with network-first fetch. Both fixes are live and verified.

Accomplishments that we are proud of

Every AI feature works end to end with real answers and honest source badges: Summarize, Extract tasks, Ask-my-notes, and a full 4-week season plan. The app saves notes offline and syncs when back in range, and the demo video shows all of it working.

What we learned

Hybrid inference is a routing problem as much as a model problem: pick the right model per request, show the source honestly, and never fake an answer. Pinning the model in config instead of code made the deprecation fix a one-line change.

What is next for Harvest Notes

On-device Nano inference in the demo once the model is downloaded in Chrome, more crops and local language support, and testing with real farmer groups.

Built With

Share this project:

Updates

Submission history