Scan a fruit leaf with your phone, get a disease diagnosis, then ask an AI agronomist for effective, natural remedies.
▶ Live Demo · Live API · Interactive Docs · Setup Guide · API Reference
The free-tier server cold-starts in ~30 s after inactivity.
| Dashboard | Scan | History | Crop Advisor |
|---|---|---|---|
 |
 |
 |
 |
Scan — Camera or photo library → EfficientNet-B0 inference → disease name, severity, and treatment list in under 1 s. Scans are stored with a signed Supabase Storage URL and linked to a plant record.
Field Map — Google Maps hybrid view with color-coded pins (green = healthy, red = diseased). Tap a pin for a slide-up panel showing the latest scan photo, confidence score, and severity.
History — Paginated scan log per plant. Plants can be renamed inline with server-side uniqueness validation.
Crop Advisor — Per-plant AI chat backed by GPT-4o-mini + Tavily web search + Serper shopping links. Builds persistent memory across conversations so the model remembers prior observations.
Drone API — Any script or drone with an API key can submit scans autonomously. GPS coordinates trigger automatic plant-record creation or matching within a 10 m Haversine radius.
EfficientNet-B0 for inference-per-cost — ~97% top-1 accuracy on the 38-class PlantVillage benchmark at 5.3 M parameters. Runs in under 200 ms on a Render free instance with no quantization.
Mahalanobis-distance OOD detection — Before scoring, the server extracts the 1280-dim backbone embedding (EfficientNet's pre-classifier layer) and computes the minimum Mahalanobis distance to the 38 per-class centroids fit on the training set. Images geometrically far from all leaf classes are rejected with HTTP 422 before consuming the user's daily scan quota — avoiding the overconfidence failure mode of plain softmax thresholding.
Keyword-gated web search — A frozenset of 60+ agricultural terms gates Tavily API calls in the chat router. Purely conversational turns skip the search entirely, cutting latency and API cost by ~70%.
Rolling 24-hour rate limiting without schema changes — POST /predict counts rows in the existing scans table filtered by user_id and created_at >= now() - 24h. No Redis, no extra columns.
API keys as SHA-256 hashes — Keys are prefixed lscan_ and generated with secrets.token_hex(32) (256 bits of entropy). Only the hash lands in the database; the raw key is shown exactly once at creation.
Per-plant AI memory — After each session, a second lightweight GPT-4o-mini call (capped at 200 tokens) extracts 1–3 observations and writes to a plant_memories table. Future sessions prepend these facts to system prompt.
Security boundary: anon key on mobile, service role on server — The mobile bundle ships only the Supabase anon key. Every DB write is enforced by Row Level Security. The service role key lives in backend env vars.
┌──────────────────────────────────────┐
│ Mobile App (Expo/RN) │
│ expo-router · Supabase anon key │
│ Row Level Security on all reads │
└────────────┬─────────────────────────┘
│ Bearer JWT
▼
┌──────────────────────────────────────┐
│ FastAPI (Docker → Render) │
│ │
│ /predict EfficientNet-B0 │
│ /chat GPT-4o-mini + search │
│ /drone/scan API-key auth, GPS │
│ /history paginated scan log │
│ /api-keys key lifecycle CRUD │
└──────┬──────────┬──────────┬─────────┘
│ │ │
▼ ▼ ▼
Supabase Tavily OpenAI / Serper
Postgres Search GPT-4o-mini +
+ Storage (grounded shopping links
(RLS) answers)