Junior computer vision and LLM engineer. I build realtime vision systems and practical LLM applications
- Watch store LLM chatbot — Dual-mode LLM chatbot with catalog-grounded product suggestion and FAQ RAG on a Qdrant vector database.
- Real-time carton detection & OCR — Live carton detection and label OCR in C++, with kg-based counting and Excel export.
- Vehicle tracking and speed estimation — Pretrained YOLOv8m, BoT-SORT, and speed between two lines. Logged run: 22.35 FPS average on a 393-frame 1280×720 clip (GTX 1660 Ti, ReID off).
- PPE detection and face recognition — Faster R-CNN for helmet, head, body, and vest, with BoT-SORT and InsightFace. Best validation mAP@0.5:0.95: 0.3957.
- Fire and smoke detection — Fine-tuned YOLOv8m, BoT-SORT, and ONNX inference. PyTorch test mAP50: 0.789.
- Lane segmentation — DeepLabV3+ on ImageNet ResNet50, exported to ONNX and run in Docker. Best validation Mean IoU: 0.7969.
- Steel defect segmentation — Custom PyTorch U-Net, trained from random initialization on Severstal. Best validation mean defect Dice: 0.60492.
C++, Python, PyTorch, TensorFlow, Keras, OpenCV, ONNX Runtime, YOLO, OCR, LLM,Lora, RAG, Prompt Engineering, Qdrant, Docker, CUDA, BoT-SORT, InsightFace.