Create README.md

README.md

@@ -1,5 +1,67 @@
 ---
 license: other
 license_name: all-rights-reserved
-license_link: LICENSE
+pipeline_tag: object-detection
+tags:
+- ppe-detection
+- yolo
+- ultralytics
+- computer-vision
+- object-detection
+- occupational-safety
+- workplace-safety
+language:
+- en
+- id
 ---
+
+# APD Vision (PPE Vision)
+
+PPE compliance monitoring from camera feeds and recordings, with reviewable evidence. APD (Alat Pelindung Diri) is the Indonesian term for PPE. Built as the final project of my AI program at rubythalib.ai.
+
+This Hugging Face page is a project card. Model weights are not published here. Source code and full documentation are on GitHub: https://github.com/arielshakaramiro/ppe-vision
+
+## What it does
+
+It monitors helmets, vests, gloves and shoes in video. Detections are linked to tracked workers, zone-specific rules are applied, and each event is saved with a photo and a short clip for supervisor review.
+
+- Live monitoring: RTSP or HTTP streams, simulated video, several cameras sharing one scheduler, automatic reconnection
+- Recordings: resumable uploads and pausable analysis jobs
+- Detection: YOLO11n-pose plus a fine-tuned PPE detector, ByteTrack tracking, and compliant, violation or uncertain states with hysteresis
+- Evidence: annotated photos, worker crops, clips, face blurring, SHA-256 hashes, supervisor review and an audit trail
+- Analytics: compliance by item, camera and zone, exposure minutes, camera quality and printable reports
+- K3 Agent: Indonesian-language Q&A over the data using read-only tools, with local Ollama or an offline router
+- Access control: viewer, supervisor and admin roles, optional Telegram alerts
+
+## Stack
+
+- Models: Ultralytics YOLO11n-pose, fine-tuned YOLO PPE detector, ByteTrack, OpenCV
+- Backend: Python 3.11, FastAPI, SQLModel, SQLite
+- Frontend: React 18, TypeScript, Vite
+- Deployment: Docker Compose with Caddy for HTTPS
+- Testing: pytest for the backend, Playwright for the frontend
+
+## Evaluation
+
+Measured on a reviewed gold set of 340 frames.
+
+| Class | mAP50-95 |
+|---|---|
+| Helmet | 0.604 |
+| Vest | 0.432 |
+
+Gloves and shoes detection is still weak and does not trigger violations by default. The four-class promotion gate has not been met, so these results are not evidence that the system is ready for field deployment. Results must be reviewed by a safety officer and are not automated safety decisions.
+
+## Datasets
+
+- SH17 (CC BY-NC-SA, non-commercial): https://github.com/ahmadmughees/SH17dataset
+- PPE Dataset for Workplace on Roboflow (CC BY 4.0): https://universe.roboflow.com/siabar/ppe-dataset-for-workplace
+- Locally annotated data
+
+## License
+
+This is a personal portfolio project and is not open source. Original code and documentation are protected, and reuse requires written permission. Third-party dependencies, models and datasets remain under their own licenses.
+
+## Author
+
+Muhammad Ariel Shakaramiro, AI Engineer (LLM applications, computer vision). LinkedIn: https://www.linkedin.com/in/muhammadariels/ and GitHub: https://github.com/arielshakaramiro