NPU graphs
Full-stack NPU inference on Qualcomm
Hand-written graphs so LLM, VLM, STT, TTS, and embeddings can run on the NPU instead of falling back to a generic runtime. Notes and kernels: contact@onchiplab.space.
Writing
Engineering notes
On-device intelligence — fast, private, hardware-native. Applied research for hardware-native AI inference. We publish claims with numbers when the run is sealed; this page is the log, not a scoreboard.
Apple GPU
Speech-to-speech on the GPU kernels
The same metal-up stack now carries speech-to-speech. We write the kernels by hand; we do not borrow a generic audio graph.
Vision
Vision language models on Apple silicon
On-device VLMs on the GPU engine. Same C++ core, same bindings, models that never have to leave the machine.