You need a vision system that works in the real world — not just on clean test data. Poor lighting, motion blur, cluttered backgrounds, occlusion, reflective surfaces. I build CV systems that are engineered for those conditions from day one, with production deployment as the design constraint, not an afterthought.

Current flagship project: A defence-grade fire and smoke detection system developed under India's iDEX DRISHTI programme, deployed on NVIDIA Jetson edge hardware for use in constrained armoured environments. Engineering constraints: early-warning latency, hard-negative rejection, sub-2% false alarm rate, and reliable inference under thermal noise and vibration. V1 achieved 74.3% mAP@50 on the D-Fire dataset. This is the kind of operating environment that forces real engineering decisions.

On the commercial side, I apply the same discipline to industrial and enterprise vision problems: safety compliance (PPE, hygiene), perimeter and intrusion detection, crowd analytics, anomaly and tamper detection, leakage monitoring, pilferage detection, and operational activity analytics.

My process covers the full cycle: use case definition, data strategy, annotation logic, synthetic data pipelines, model development, error analysis, false positive reduction, and edge-oriented deployment.

Stack: Python · OpenCV · PyTorch · YOLOv8/YOLO11 · ONNX · TensorRT · NVIDIA Jetson (Orin series) · Roboflow · Ultralytics
If you are building a vision system that needs to work in a hard environment — I am interested in the problem.