AgriXpert case study: a farming app for smallholder farmers

01 — The problem

Good farming advice is scattered, and rarely in your own language.

A smallholder farmer who spots holes in their maize leaves has to find an extension officer, ask around, or search online in English. Weather, input prices, buyers and casual workers are each found somewhere else again.

The brief was to put expert farming advice, local weather and a place to buy, sell and hire into one app. It had to:

Answer everyday crop, livestock and pest questions, with an AI assistant for anything the guides don't cover
Work in English, Shona and Swahili
Keep notes and farm tools working offline
Connect farmers with buyers, sellers and workers
Keep running costs low, so the app isn't priced out of reach

02 — How we built it

The same six steps we use for every client.

AgriXpert has had two versions so far. Here's what each step looked like on this project.

1

Brief

One app for advice, weather and trade, usable by farmers in their own language. We also set a hard rule: everything runs on free service tiers, with no paid cloud features.
2

Design decisions

We split the app into product areas (crops, livestock, AI support, weather, community, marketplace, jobs) and decided what lives on the phone and what lives online. Notes and tools stay on the device; posts, orders and jobs go to the cloud. More on that below.
3

Clickable prototype

We mapped the screens a farmer uses most (home and weather, the assistant, pests and diseases, the fertilizer calculator, community, market and jobs) as a clickable prototype, so the flow can be checked on a phone. Try it below.
4

Build in small pieces

The first version covered the crop and livestock tools, the AI chat, weather, a community feed and a marketplace. The current version added the farm job board, the three languages, a responsive layout for tablets and desktops, and a full clean-up of the code into one folder per product area.
5

Test and harden

Reviewing the first version, we replaced its home-made login with Firebase Authentication, took the AI and weather keys out of the app, and fixed bugs such as the chat forgetting earlier messages. Automated tests now cover the fertilizer maths, the assistant's topic filter, image compression, the cart and the data models, and they run on every push.
6

Ship and support

Not released yet. The app has been run on an Android emulator, where the main screens open without crashes, and the remaining launch work (connecting the live Firebase project, deploying the weather proxy and database rules) is written down in the repo.

03 — Try it

Tap through the prototype.

This is the clickable prototype, running live on the page. It uses sample names, prices and weather, and it may be a step behind the real app.

Switch the language between EN, SN and SW at the top of Home
Open Assistant and tap “My maize leaves have holes”
Try the fertilizer calculator, or open Market and post a farm job

Open full screen ↗

Real screens from the first version (2025)

Screenshots of the app running on an Android phone. Scroll sideways to see them all.

Welcome screen: smart farming at your fingertips — Welcome

Home screen with local weather, services, notes and planned activities — Home

AI support: chat, disease scanner, recommendations and alerts — AI support

Community feed with farmers sharing crop photos — Community

Weather screen with farming insights — Weather

Five-day forecast with advice for each day — 5-day forecast

04 — Key decisions

Four choices that keep it cheap and safe to run.

In plain English, without the jargon.

Free tiers only

Accounts, posts, orders and jobs run on Firebase's free plan. That plan has no file storage, so photos in community posts are shrunk on the phone and saved inside the post itself, kept well under the size limit.

No secret keys inside the app

Anything built into an app can be dug out of it. So the AI assistant goes through Firebase, which needs no key in the app, and weather goes through a small free proxy that holds the weather key and caches answers for about ten minutes per area.

An assistant that sticks to farming

The AI is told to answer only farming questions, keep it short and practical, and suggest a local agronomist or vet for serious problems. Off-topic questions get a polite “I can only help with farming” before they reach the AI.

Offline where it matters

Notes, planned activities and farm tools are saved on the device, so they work without signal. If the cloud isn't set up or can't be reached, online features show a clear “not available” message instead of crashing.

The plant disease scanner runs a small image model on the phone itself, so a photo of a leaf can be checked without uploading it. Payments are kept simple for now: orders use cash on delivery or mobile money, confirmed by hand, and the app collects no card details.

05 — Tech stack

What it's made of.

App

Flutter for Android, iOS, web and desktop; Provider and Bloc for state; bottom bar, rail or sidebar by screen size

Cloud data

Firebase Authentication (email and Google) and Cloud Firestore, with access rules for posts, comments and profiles

On the device

SQLite and shared preferences for notes, activities and tools; works offline

AI

Gemini through Firebase AI Logic for the assistant; an on-device TensorFlow Lite model for leaf disease detection

Weather

OpenWeather through a Cloudflare Worker proxy, using the phone's location

Languages & quality

English, Shona and Swahili translation files; automated tests and code checks on every push

06 — What's next

Still being built.

AgriXpert is in development and not released yet. Before launch we need to connect the live cloud project and deploy the weather proxy. After that, these are on the list but not built yet:

Native speakers checking the Shona and Swahili wording
Push notifications
A real payment provider for marketplace orders
Saving changes made offline and sending them when signal returns
An accessibility review

Farming advice, weather and a market
in one app, in the farmer's language.

Good farming advice is scattered, and rarely in your own language.

The same six steps we use for every client.

Brief

Design decisions

Clickable prototype

Build in small pieces

Test and harden

Ship and support