At ThoughtClan Technologies, I built a real-time re-identification system for multi-camera RTSP CCTV. It assigns visitors persistent IDs across cameras and achieved 0.97 AUC and 85% TPR at 1% FPR.
I also evaluated a YOLO11n helmet and PPE detector, which achieved 93.1% mAP@0.5 on a held-out test set. On a Raspberry Pi 5, I benchmarked CPU-only edge inference and measured 5.61 FPS end-to-end throughput.
At AI4SEES, I adapted a YOLOv5-6D-Pose model from X-ray to RGB inputs for 6DoF object pose estimation. The model achieved 6° rotation error and 3 cm translation error on a custom synthetic dataset.
For my Diffender project, I built a diffusion-based defense that localizes and restores adversarial patches without retraining the frozen diffusion backbone. I reduced inference latency by 38% while achieving 85.7% robust accuracy against AdvP patch attacks.
