Detection • Tracking • Re-ID • 3D • Optimization (TensorRT / TFLite). I design, train, and deploy systems that run on edge and cloud.
Edge-optimized tracker (TensorRT) at 60 FPS on Jetson Orin. Robust to occlusion and low light.
Person re-identification with privacy-aware embeddings and heatmaps for dwell/time-in-zone.
Keypoint models + IK for grasp planning; sub-50 ms end-to-end with quantization and pruning.
Practical tricks for MOT stability: motion priors, re-ID hints, and temporal smoothing.
Quantization, TensorRT/TFLite paths, and what to profile before choosing hardware.
Input drift, alerting, and dataset refresh: an MLOps checklist for vision systems.
Fast experiments, success metrics, and an MVP you can demo to stakeholders.
Torch → TensorRT/TFLite, INT8 calibration, latency/throughput trade-offs.
APIs, streaming pipelines, monitoring, retraining loops, and on-device updates.