Chitr | Real-Time Loss Prevention Platform

Professional project showcase for Chitr, a production-grade retail loss prevention system built for real-time CCTV analysis.

This repository intentionally contains project documentation only. No source code, model weights, credentials, runtime configs, or deployable internals are published here.

Executive Summary

Chitr is a local-first AI platform that monitors live RTSP camera streams and raises guard alerts for suspicious concealment behavior.

Core objective:

• reduce shrinkage and investigation lag
• keep false alarms low enough for sustained daily usage
• operate with no cloud dependency during runtime

Measured Outcomes

• ROC-AUC: 0.941
• Accuracy: 93.3%
• False alarm rate: 2.7%
• Fast-path gate latency: under 500 ms (pose stage)

System Design

Three-stage alerting pipeline:

1. Motion detection gate
2. YOLOv8 pose suspicion scoring
3. Deep clip scoring (VideoMAE + SlowFast ensemble)

Alert output:

• Telegram snapshot alert
• follow-up video clip
• guard feedback signal for retraining

For architecture detail, see ARCHITECTURE.md.

My Role

• designed and iterated the multi-stage inference pipeline
• integrated FastAPI backend with live camera orchestration
• implemented alerting + human feedback loop
• connected model lifecycle metrics to MLflow tracking
• optimized deployment for local edge constraints

Technology Stack

• Python
• PyTorch
• YOLOv8
• VideoMAE
• SlowFast R101
• FastAPI
• PostgreSQL
• OpenCV
• MLflow

Security and IP Policy

This is a showcase repository only.

• no runtime secrets are published
• no source code is published
• no reusable model artifacts are published
• all implementation IP remains private

For policy detail, see SECURITY_AND_ACCESS.md.

Contact

Priyansh Patel

• LinkedIn: https://www.linkedin.com/in/priyansh-1221/
• GitHub: https://github.com/priyansh1221