At OmniClouds, I wrote a custom C++ video inference pipeline using hardware video decode, CUDA, and TensorRT. It replaced a slower OpenCV and PyTorch YOLO setup, delivering about 1.6x the throughput with a third of the system RAM.
I built facial recognition and real-time number plate recognition modules on the GPU pipeline, and added people-counting and line-crossing analytics using multi-object tracking. I also trained an object detection model for PPE violations, reaching 86.8% mAP50.
For on-premises clients, I set up deployment components including offline licensing, isolated tunnels, client authentication, and a Docker stack. Before that, I designed and deployed enterprise SD-WAN setups at OmniClouds.
In personal projects, I built Annotator, a browser-based image labelling tool, and Auto Alert Handler, an LLM agent that finds runbooks with RAG and either acts on infrastructure alerts or escalates them to a human.
