At Sergek Group, I build production video analytics across face recognition, anti-spoofing, and vision-language models. I work across the full cycle, from data and training to optimised CPU and GPU deployment.
I built a two-stage driver seatbelt detection pipeline using YOLO-World and SAM3 to localise the driver and filter frames at the edge. Verified triggers go to a Qwen-VL model served with vLLM, where visual reasoning helps keep false positives low.
At Sergek Group, I also designed and shipped a multi-task face analysis model with six attribute heads trained on shared ArcFace features. I merged the heads with the backbone into a single ONNX graph that returns identity embeddings and attributes in one forward pass.
I trained an anti-spoofing model and integrated it into the production face recognition pipeline as a gate before identity matching. Earlier, at TargetAI Limited, I co-developed “Backyard,” an internal tool for the support team; in university projects, I compared classical models with LSTM-based sales prediction and trained a TCN model for motor fault detection.
