Computer vision and edge deployment: model optimisation, C++ inference runtimes, and the service around them. I am most useful when a model has to run on real hardware with a latency budget.

Qualcomm QNN (HTP) on Snapdragon, TensorRT and ONNX Runtime on NVIDIA, ExecuTorch and TFLite on mobile, llama.cpp on CPU, plus CUDA, Vulkan and Hexagon device selection.

Usually yes. Profiling first: find where the time and memory actually go, then decide between graph changes, quantisation, a different runtime, or a smaller architecture.

I am open to conversations about Edge AI, inference runtimes and AI products. Email is the fastest way to reach me; I reply within one business day.

Send the model or task, the target hardware, and what success looks like. A repository, a benchmark number or a failing profile is even better.