Toan T.

I build real-time computer vision and Edge AI systems, mainly on NVIDIA Jetson, using C++/Qt and GPU inference. Most of my work is around making vision systems run reliably and efficiently on real hardware. Computer Vision / Edge AI * Retail shelf recognition on Jetson Orin NX using YOLO, TensorRT/CUDA and ByteTrack. The pipeline runs at around 13 ms per frame, with GPU NMS taking less than 1 ms. * Product recognition with unknown-product rejection and live switching of recognition models without restarting the video stream. * Person, fall and accident detection, including tracking through detector or pose failures using pixel-level evidence. * Open-vocabulary detection where the class list can be changed per camera at runtime without retraining. C++ / Qt * C++ and Qt/QML applications, including legacy code, migration, debugging and performance optimization. GPU / LLM Inference * Working with vLLM and llama.cpp to understand and benchmark real inference workloads. * Looking at the parts that directly affect performance: prefill vs. decode, KV cache, batching, GPU utilization and kernel-level behavior.