🍎 OrchardEye
A dual-sensor apple health & quality scanner for Washington's 8th Congressional District.
One scan of a fruit or leaf → two answers:
- Is it diseased? — on-device computer vision classifies visible disease (or "healthy").
- Is it good? — a low-cost clip-on near-infrared (NIR) sensor estimates ripeness and sugar (Brix) — internal quality the camera can't see.
…then fuses both into one plain-language verdict, speaks it aloud (weather-aware voice guidance), and tracks each plant's trend over the season — all on sub-$130 hardware that works offline in the field.
Built for the Congressional App Challenge — WA-08 (tree-fruit country: Wenatchee, Chelan, Leavenworth). One crop, done properly: apple, the district's and the state's defining tree fruit.
Pitch: "OrchardEye puts a $3,000 lab capability into a $100 phone attachment so the small apple growers who power my district can pick at peak quality and catch disease early — turning waste into income."
⚠️ Status: v0.1. The planning artifacts (PRD, research, designs, build prompt) are in docs/, and a buildable native-iOS app is in OrchardEye/ — full UX with the disease model and BLE sensor stubbed behind protocols (see the app README).
📱 The app
A complete native iOS (SwiftUI) v1 lives in OrchardEye/ — it builds in
Xcode 16+ and runs the full scan → quality gate → disease + crop-health → NIR ripeness/Brix →
fused verdict → voice + trends + confirm/correct flow, all on-device.
open OrchardEye/OrchardEye.xcodeproj # ⌘R on an iPhone simulator
It's Simulator-friendly: tap a sample leaf and Scan fruit (the camera/NIR are simulated)
to walk the whole pipeline. The disease model and BLE sensor are clean stubs behind protocols
(DiseaseClassifier, SpectralSensor) ready for a real Core ML model and an AS7265x. See
OrchardEye/README.md for build notes and the real-vs-stubbed map.
🌐 Landing page
A self-contained marketing site lives in web/. Plain HTML/CSS/JS, with a
tiny root build script for Vercel that copies web/ to dist/. Same brand palette
as the app (orchard green + ripe amber).
cd web && python3 -m http.server 8000 # → http://localhost:8000
Build/deploy notes for Vercel and GitHub Pages: web/README.md.
📚 Documentation index
| Doc | What's inside |
|---|
| 🌐 Landing page | Self-contained marketing site (hero → problem → how-it-works → sensors → pilot). Deploys to Vercel or GitHub Pages. |
| 📱 iOS app | The buildable SwiftUI v1 (Xcode 16+): full scan→verdict→trends flow, on-device. |
| PRD | Full product requirements: goals, users, features, architecture, BOM, ML plan, milestones, risks, CAC alignment. |
| AI Pipeline | The end-to-end flow: quality gate → preprocess → quantized on-device models → fusion → report/trends → community loop (with a Mermaid flowchart + feedback loops). |
| Feature List & Variables | The four core variables (disease, crop-health score, trend/risk detection, voice guidance) + supporting features + the variable dictionary. |
| Community & Voice Vision | The "matured" layer: weather-adaptive voice guidance, farmer-wisdom corpus, and community-owned localized AI. |
| v1 Build Prompt | A comprehensive, copy-paste prompt (master + phase prompts) to drive Claude/ChatGPT while building v1. |
| Research · Disease Loss | Apple-loss-to-disease statistics across US, WA, and WA-08. |
| Research · Competitors | What already exists: disease-ID apps, NIR/Brix tools, food-spoilage apps, retail grading — and where OrchardEye fits. |
🧠 How it works (30-second version)
📷 camera ─┐ ┌─→ 🦠 disease + crop-health score
├─→ quality gate → preprocess → quantized on-device models → 🧠 fusion ─→ 🟢/🟠/🔴 verdict + 🔊 voice
🔬 NIR ────┘ (offline, < 3 s) └─→ 🍎 ripeness + Brix + 📈 trend chart
Full detail with feedback loops: docs/design/ai-pipeline.md.
🔩 Hardware (~$118 BOM)
AS7265x spectral sensor (18 ch, 410–940 nm) + ESP32 (BLE bridge) + Brix refractometer (ground-truth) + 3D-printed dark shroud. Full list in PRD §8.
Honest-claims rule: the AS7265x tops out near 940 nm — great for ripeness, chlorophyll, surface chemistry. We do not claim deep sugar/water bands (1450 nm), which need far pricier sensors. Rigor is the point.
✅ v1 success targets
| Target |
|---|
| Disease accuracy (public test) | ≥ 90% |
| Disease accuracy (own field photos) | ≥ 75% (honest domain gap) |
| Ripeness classification | ≥ 85% |
| Brix regression | R² ≥ 0.80, RMSE ≈ 1 °Brix |
| Capture → verdict | < 3 s, on-device |
| Hardware BOM | < $130 |
🗺️ Roadmap
Apple disease MVP → sensor/BLE bridge → NIR calibration set + quality models → fusion + voice + trends → field test → demo & submission. See the milestone table.
Working name "OrchardEye" is a placeholder. Verify the current CAC deadlines, district boundaries, and representative before submission.
