MangoGuard — Edge AI That Detects Mango Diseases in the Field

Eyobel — Wed, 03 Jun 2026 03:53:06 +0000

Submission for the GitHub Finish-Up-A-Thon Challenge

Mango farming is a lifeline for millions of smallholder farmers in Ethiopia.
A single fungal outbreak can silently destroy 20–30% of a harvest before a
farmer even recognises it. Existing solutions require a lab, a specialist,
or reliable internet. None of those exist where the problem is worst.

MangoGuard runs AI directly on a microcontroller smaller than a credit card
— no cloud, no Wi-Fi, no lab.

An Arduino Nano 33 BLE Sense runs a quantized MobileNetV1 model that
classifies mango leaf disease in under 2 seconds at 86.45% accuracy on real
Ethiopian farm data. The Nano also reads live temperature and humidity via a
DHT22 sensor and sends everything — disease result, temp, and humidity — to a
Raspberry Pi 4 gateway, which evaluates environmental disease risk against
agronomic thresholds, runs a 24-hour AI forecast model, and generates
plain-language recommendations pushed directly to farmers based on current
conditions.

Everything streams to a React dashboard in real time via WebSocket.

From the dashboard you can:

Monitor live environmental readings and disease risk
Run a cloud scan from your phone — upload any leaf photo for instant AI classification, no hardware required
View a 5-day disease risk forecast calendar
Browse your full scan history in the logs section

The dashboard is fully bilingual — English and Amharic (አማርኛ) — because
agronomists advising Ethiopian farmers shouldn't have to work in a language
that isn't theirs.

Under the hood, every scan is saved to a PostgreSQL database. The admin
dashboard exposes this data as a labeled dataset that can be used to
retrain and improve the model over time as more field data comes in.

🌐 Live Demo: https://mango-guard.vercel.app/
📂 GitHub: https://github.com/SCIFI-Shinobi/Intelligent-Mango-Health-Monitoring

The Comeback Story

The hackathon prototype worked — but only on my machine, with undocumented
secrets, and no way for anyone else to run it. Here's what I shipped to fix that:

Rewrote the README from scratch — setup guide, env variable reference, troubleshooting for the 7 most common failures
Added ARCHITECTURE.md, DEPLOYMENT.md, and CONTRIBUTING.md
Fixed .gitignore which was blocking .env.example files from being committed
Documented all 10 required environment variables (previously scattered in source code)
Added GPL-3.0 LICENSE — the repo had no legal clarity before
Fixed a real production bug — I had set scan intervals to 20 seconds and forecast to trigger after 3 readings for demo convenience. The forecast model was actually trained on a 24-reading window (one per hour). In the field it was producing numbers that looked valid but weren't. Fixed: scanIntervalMs = 3600000UL, forecast threshold ≥ 24.

A stranger can now fork, configure, and deploy this in under 20 minutes.

Copilot Experience

Copilot saved the most time on the backend — main.py grew past 3,000 lines
and it was excellent at continuing repetitive patterns: email templates,
database migration helpers, route structure. Once I wrote the first, it nailed
the second.

During the polish phase, asking Copilot to review the README flagged that I
had no troubleshooting section and no env variable docs — exactly what a
first-time contributor needs. It also helped clean up firmware comments after
I fixed the production bug.

It spotted multiple ESLint warnings, duplicate translation keys, and unused imports/variables that were silently compiling locally but completely blocking production builds on Vercel CI.

The honest limitation: it continues patterns well but won't proactively tell
you what's missing. You have to ask the right question first.

Built to bridge the gap between AI and smallholder agriculture in Ethiopia. 🌱

DEV Community: Eyobel

MangoGuard — Edge AI That Detects Mango Diseases in the Field

The Comeback Story

Copilot Experience