Smart insect traps: the camera shoots, the AI counts.
FlyBot watches sticky and lure traps through EZVIZ cameras, counts the insects on every photo with AI (fruit flies first), and tells growers when a trap crosses its threshold. It runs in the browser and installs on the phone like an app, in Vietnamese and English.
Website: https://ezviz.agrhub.com · Contact: contact@agrhub.com
Commercial (1080p, 60 s, with music): English · Tiếng Việt; 9:16 for social media (30 s): English · Tiếng Việt. Product tour (1080p, 86 s): English · Tiếng Việt. The files are also in the repository: client/public/media.
- Catches per trap per day. Sticky traps count the daily rise on the glue card; lure traps count the average per photo. Flies per trap per day compare traps and weeks fairly.
- AI detection with an RF-DETR model trained on Roboflow (camos/fruit-fly-wmcqi, 96% mAP@50 on the test split; insect boxes and species) and Google Vertex AI (Gemini) for diagnosis and photo review. People can confirm or correct each count; corrected photos can go back to Roboflow for retraining.
- Cameras: live view, SD-card playback on a timeline, scheduled captures, image-quality checks and a card-replacement reminder.
- Alerts by email, Telegram and push notification, with per-trap thresholds, quiet hours and a 06:30 morning brief.
- Analytics and reports: charts per trap, hour and period, a 7-day forecast and degree-day outlook, report builder with scheduled PDF emails, shareable links with expiry, PDF/Excel export with the detection boxes.
- FlyBot assistant: an agent that looks up real data (catches, alerts, cameras, reports, treatments), shows photos and charts, takes a new photo, and proposes changes that only apply after you confirm them.
- Farms and organizations: farms with plots drawn on a map, a treatment logbook with VietGAP export, organizations with owner/manager/member roles, and sharing per camera, farm or report.
- Admin: users and roles, all cameras, audit log, system settings and a capture pipeline monitor.
- Public API and webhooks for other systems (API.md).
 Overview: flies caught today, flies per trap per day, every trap against its threshold |
 AI detection: every fly boxed and counted; confirm or correct the count |
 FlyBot assistant: answers from real data, with photos and charts |
 In-depth analysis: per-trap table, daily catches, temperature, catches by hour |
 Reports: build, export to PDF, share with expiry, send weekly or monthly |
 Field map: traps, plots and a density heat map |
 Cameras: live view and SD-card playback on a timeline |
 Field log: treatments with their effect on catches, VietGAP export |
Installs on the phone like an app, with push notifications
Screenshots in Vietnamese are in client/public/manual/vi.
| Part | Built with |
|---|---|
Web app (client/) |
Vue 3, Vite, Element Plus, vue-i18n, Chart.js, Leaflet |
API server (server/) |
Node.js, Express 5, Mongoose 9 (MongoDB), node-cron |
| Storage | S3-compatible object storage for photos |
| AI | Roboflow workflows, Google Vertex AI (@google/genai) |
| Hosting | Docker: Google Cloud Run or a Compute Engine VM (DEPLOY.md) |
client/ Vue web app (pages, components, locales vi/en, guide screenshots in public/manual)
tests/ Vitest unit and component tests, end-to-end smoke test on the demo
server/ Express API, cron jobs (capture, AI retry, reports, cleanup), assistant agent
src/agent/ FlyBot assistant: prompts, tools, Gemini streaming loop
tests/ node:test suite with an in-memory MongoDB
scripts/ maintenance scripts (rebuild stats, find orphan files, VAPID keys...)
deploy/ Cloud Run deploy script and settings template
docs/ AI model and dataset, README images
tools/media/ demo server, screenshot and product-tour recording tools
tools/perf/ performance benchmarks
API.md public API and webhooks
DEPLOY.md Cloud Run and VM deployment
Requirements: Node.js 22.12+, a MongoDB database, S3-compatible storage, an EZVIZ Open Platform app (app key and secret) and a Roboflow workflow.
The product tour videos and the presentation decks are stored with Git LFS: run git lfs install once before cloning (or git lfs pull afterwards), or those files are small pointers.
npm install
cp .env.example .env # then fill in the values (see the comments in the file)
npm run dev # API on :3003, web app on :3000The first account that signs up becomes the admin. JWT_SECRET must be at least 32 random characters, or the server refuses to start.
Optional: SMTP settings for emails, TELEGRAM_BOT_TOKEN for Telegram alerts, VAPID keys for push notifications (node server/scripts/generate_vapid_keys.js), VITE_GOOGLE_MAP_KEY for Google Maps layers, and a GCP service account (GOOGLE_APPLICATION_CREDENTIALS) for the assistant locally. See server/README.md for roles, push notifications and the server settings.
| Command | What it does |
|---|---|
npm run dev |
API and web app with reload |
npm test |
Server and web app tests (in-memory MongoDB, no external services) |
npm run test:e2e |
The demo in headless Chrome: every main page and the main flows |
npm run build |
Web app into client/dist, server bundle into server/dist |
npm start |
Production server from server/dist |
node deploy/deploy-cloudrun.js --check # what .env provides for production, no gcloud
deploy/deploy-cloudrun.sh # or deploy\deploy-cloudrun.bat on WindowsThe script sets up the Google Cloud project, stores secrets in Secret Manager, builds and deploys to Cloud Run, then checks /health. For a VM, use docker compose up -d --build. Everything is in DEPLOY.md.
- DEPLOY.md: deployment, settings and post-deploy checks
- API.md: public API and webhooks
- docs/AI_MODEL.md: the fruit fly detection model, its dataset and scores, and how to retrain it
- server/README.md: server configuration, roles, push notifications
- client/UI_GUIDELINES.md: UI rules for the web app
- tools/media/README.md: demo, screenshots and product-tour video
- CONTRIBUTING.md: development conventions and tests
- SECURITY.md: reporting a vulnerability
- CHANGELOG.md: what changed
- In the app: the Guide page (
/manual), with screenshots in both languages
AGPL-3.0 © 2026 AGRHUB. Versions up to 1.5.0 were released under the MIT licence.