| Project | Stack | Description |
|---|---|---|
| yolo-pose-cpp | C++, ONNX Runtime, CUDA, Docker | GPU-oriented pose estimation pipeline with zero-copy binding. Solved CPU preprocessing bottleneck (~63% time). 640×640: 37→148 FPS (+300%), 1280×1280: 21→58 FPS (+176%). |
| yolo-trt-nms | Python, TensorRT, Triton, PyCUDA | One-command YOLO → TensorRT export with GPU-accelerated NMS. NMS baked into graph via EfficientNMS_TRT. Auto-generated Triton config. Threading pipeline + dynamic batching out of the box. |
| YOLO vs RT-DETR Benchmark | PyTorch, ONNX, TensorRT | Comparative benchmark on RTX 5080 across inference frameworks. Performance analysis for YOLO and RT-DETR models with different export formats. |
| Retail Analytics Platform | PostgreSQL, Airflow, PySpark, FastAPI | End-to-end system: parsing, ETL pipelines, analytics. 73M+ orders processed. Telegram bot with analytical reports. |