README.md

metadata

title: ReefScan API
emoji: 🪸
colorFrom: green
colorTo: blue
sdk: docker
app_port: 7860
pinned: false

ReefScan backend (FastAPI, async-job inference)

CPU inference API for coral-health analysis: SAM2 segmentation → DINOv2 classification → conformal prediction sets. Loads weights + calibration from HrishiKabra/reefscan-dinov2-coral.

Endpoints

• POST /infer — multipart file (image/video) or form url → 202 {job_id}
• GET /infer/{job_id} — poll → InferenceResponse (status + results)
• GET /health — model load status + last latency
• GET /review-queue, POST /review-queue/{id}/confirm
• GET /reef-locations, GET /reef-locations/{id}/snapshots
• GET /observability

Space secrets (Settings → Variables and secrets)

name	required	purpose
SUPABASE_URL / SUPABASE_ANON_KEY	for logging	inference_logs / review_queue / snapshots / jobs
R2_ENDPOINT / R2_ACCESS_KEY_ID / R2_SECRET_ACCESS_KEY / R2_BUCKET	for image display	store uploaded images
HF_MODEL_STAGE	optional	linear_probe (default) or finetune
HF_TOKEN	only if model repo is private	weights download
REEFSCAN_STUB	optional	1 forces synthetic inference (no models)

Without Supabase/R2 the API still runs: logging is a no-op and image urls are placeholders. First request is slow (cold start downloads ~0.7 GB of weights); SAM2 AMG is ~15–25 s/image on the free 2-vCPU box (Phase 1.5), which is why inference is an async job.