Self-taught Linux engineer. I deploy Whisper ASR and LLMs on AMD Ryzen AI NPU hardware — end to end, from the kernel driver stack to an OpenAI-compatible API. Hands-on, hardware-near, and fully local.
The full journey: amdxdna/XRT driver bring-up, the one real gotcha (memlock), FastFlowLM instead of the Ryzen AI SDK, and a measured NPU-vs-CPU energy comparison — whisper-large-v3-turbo at RTF ≈ 0.18 for roughly a tenth of the energy.