Independent R&D
Making computation fast on real chips.
AI models · GPU code · small devices · scientific simulation. Measured on hardware I own, with numbers you can verify.
Scope
Research tracks
My R&D asks one question: how does computation really run on real chips? I work on it through AI inference, GPU programming, small devices and scientific computing.
Latest result
Where does each lock hold?
Three ways of locking a language model compared on one Mac: safety training, a system-prompt instruction, and an outside filter.
Safety training wrongly stopped 2 of 100 harmless questions. Adding Llama Guard raised that to 21 of 100. A plain-text secret leaked in 45% of tricks on Llama; a simple code check cut leaks to 7.5%. Five runs each, 980 answers, on one Mac.
Field notes
Latest notebook
Where does each lock on a language model hold, and what does it cost?
Three ways of locking a language model compared on one Mac: safety training, a system-prompt instruction, and an outside filter.
A reviewer's question: how much of the speed-up is batching?
A batched Transformers baseline shows most of the earlier speed-up was batching; at identical work, vLLM was about 1.7 times faster.
Why repeated identical prompts look faster: prefix caching
Re-sending the same prompt can look much faster because engines reuse computed KV for a shared prefix. That is a real effect — and a trap if you ignore it in a benchmark.
Contact