Acre — Handheld Plant Scanner

A field-deployable, offline-first AI plant scanner for the AI Hackathon @ Berkeley 2026.

You carry Acre and point it at a plant. On-device, it identifies the plant, detects disease and pests, and computes a 0-100 health score — then lights an LED on the Raspberry Pi: green = healthy, red = needs treatment. The red LED is the spray-substitute ("treat this one"). Everything intelligent runs locally on the Pi 5 / QNX board; the cloud is reporting-only (map, pesticide list, AI summary).

See the full spec in docs/Acre_PRD.md.

QNX Pi + laptop demo (friend's setup): docs/DEMO_RUNBOOK.md

Architecture: local does the thinking, cloud does the reporting

flowchart LR
  subgraph edge [On device: Pi 5 + QNX, fully offline]
    cam[Pi Camera Module 3] --> det[YOLOv8n + disease/pest classifiers, ONNX]
    det --> score[Per-plant health score]
    score --> ledNode[Red/Green LED on Pi GPIO]
    score --> db[(Local SQLite, RTC-stamped)]
  end
  db -->|opportunistic sync| apiNode[Cloud API]
  subgraph cloud [Cloud: reporting only, no inference]
    apiNode --> pg[(Postgres / SQLite)]
    pg --> webNode[Web map + report]
    pg --> mon[Arize / Poke]
  end

Repo layout

Path	What it is
docs/Acre_PRD.md	Product requirements document
edge/	On-device pipeline: capture, ONNX detection, classifiers, ArUco zones, health score, GPIO LED, SQLite, sync agent
models/	Off-device training + ONNX export for the 3 models
cloud/	FastAPI reporting backend (sync ingest, aggregation, Claude summary)
web/	Next.js dashboard: farm map, pesticide table, AI summary
cloud/integrations/	Optional Arize (monitoring) + Poke (conversational report)

Quickstart

1. Edge (runs on a laptop too, with stubs)

pip install -r requirements.txt
python -m edge.main --once          # single scan: prints finding + LED state
python -m edge.main                 # continuous handheld scan loop
python scripts/webcam_farm_demo.py    # webcam + press F → live health report in browser

On a dev laptop with no camera/models it uses synthetic frames, disables missing model stages, and prints LED states. On the Pi it lights the real LED via GPIO. Set ACRE_SENSORS_ENABLED=1 to include the optional environmental sensors.

2. Cloud (reporting API)

pip install -r cloud/requirements.txt
python -m cloud.seed                                  # farm + zones + UC IPM + demo data
uvicorn cloud.app.main:app --reload --port 8000

Point the device at it: ACRE_BACKEND_URL=http://<host>:8000/api/sync. Defaults to local SQLite; set ACRE_DATABASE_URL for Postgres/Supabase.

3. Web dashboard

cd web
npm install
cp .env.example .env.local
npm run dev                          # http://localhost:3000

4. Train the models (off-device)

cd models
pip install -r requirements-train.txt
python train_detector.py --data data/weed_crop/data.yaml --epochs 50
python train_disease_classifier.py --data data/plantvillage --epochs 15
python train_pest_classifier.py --data data/ip102_subset --epochs 15
# -> artifacts/*.onnx copied to the Pi; edge/detect.py loads them

Key environment variables

Var	Default	Used by
ACRE_BACKEND_URL	http://localhost:8000/api/sync	edge sync agent
ACRE_HEALTH_THRESHOLD	70	edge LED green/red cutoff
ACRE_SENSORS_ENABLED	0	edge optional sensors
ACRE_DATABASE_URL	sqlite:///./acre_cloud.db	cloud DB
ANTHROPIC_API_KEY	(unset → offline summary)	cloud AI summary
NEXT_PUBLIC_ACRE_API	http://localhost:8000	web app

Hardware (handheld build)

Pi 5 (QNX) + Pi Camera Module 3 (CSI) + RGB LED & buzzer & 1602 LCD on Pi GPIO + DS1302 RTC. No servo, no laser, no Arduino on the critical path — the LED replaces the laser/spray. Full inventory and wiring rationale in the PRD (sections 5-7).