CoolFace
Apppublic

18-monarch/credisafe-vision

sourceHugging Faceupdated 3mo agoView on Hugging Face
0likes
App README

CrediSafe Vision Service

The production-facing Python service for CrediSafe. It combines plate OCR with review-only road-scene observations and returns structured evidence to the Next.js application.

Verified capabilities

  • —Samples short uploaded road or vehicle clips with OpenCV.
  • —Locates plate-shaped regions and reads Indian registration numbers with multipass Tesseract OCR.
  • —Combines repeated reads across frames and compares the result with the registered vehicle.
  • —Uses the bundled models/yolov8n.pt COCO model for general objects such as vehicles, people, traffic lights and phones.
  • —Produces review evidence and transparent confidence values.
  • —Deletes the temporary uploaded file after every request.

Deliberate boundaries

  • —Phone proximity and traffic-light colour are observations for human review, not automatic violations.
  • —Red-light compliance, wrong-side driving and lane discipline require calibrated camera geometry and are not inferred from arbitrary uploads.
  • —Helmet detection remains disabled until a validated dedicated helmet model is configured.
  • —This service never awards XP, removes XP or issues penalties. CrediSafe's GPS and database settlement logic remains authoritative.

API

MethodEndpointPurposeAuthentication
GET/Service metadataPublic
GET/healthRuntime and capability healthPublic
GET/docsOpenAPI documentationPublic
POST/v1/analyzeAnalyse a short video clipX-CrediSafe-Key

POST /v1/analyze accepts multipart form fields:

  • —video — required video file
  • —expected_plate — optional registered number
  • —trip_id — optional CrediSafe trip identifier

Required hosted secret

Add this in the hosting platform's secret manager:

env
VISION_API_KEY=use-a-long-random-production-secret

Use the exact same value in Vercel as a server-only variable. Never prefix it with NEXT_PUBLIC_.

Recommended Hugging Face Space variables

env
VISION_ENV=production
VISION_REQUIRE_API_KEY=true
VISION_ALLOWED_ORIGINS=https://your-vercel-domain.vercel.app
VISION_MAX_UPLOAD_MB=25
VISION_SAMPLE_INTERVAL_SECONDS=0.4
VISION_MAX_SAMPLED_FRAMES=120
VISION_MAX_CONCURRENT_ANALYSES=1
VISION_QUEUE_WAIT_SECONDS=3
VISION_INCLUDE_EVIDENCE_IMAGES=true
VISION_PRELOAD_MODELS=true
VISION_EXPOSE_INTERNAL_ERRORS=false
VISION_MAX_CANDIDATES_PER_FRAME=3
VISION_MIN_FRAME_SHARPNESS=18
VISION_OCR_TIMEOUT_SECONDS=2.5

The default bundled object-model path is resolved automatically. Do not set VISION_OBJECT_MODEL_PATH unless replacing the bundled model.

Local development from the monorepo

From the CrediSafe repository root:

powershell
py -3.12 -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -r vision_service\requirements.txt
$env:VISION_API_KEY="credisafe-local-secret"
npm run dev:vision

Health check:

text
http://127.0.0.1:8000/health

Docker from the backend folder

powershell
docker build -t credisafe-vision .\vision_service
docker run --rm -p 7860:7860 `
  -e VISION_API_KEY="replace-with-a-long-secret" `
  credisafe-vision

Persistence

The container filesystem is temporary. Videos are processed and deleted locally; users, trips, XP, rewards and analysis records belong in Supabase.