How Penwright stitches whisper.cpp and llama.cpp together
A short technical tour of the on-device AI pipeline.
DraftingFresh notes
Voice first. Notes follow.
Essays on local AI, Mac tooling, and the math of typing vs. speaking. New posts land when we have something useful to say.
Published
Engineering·8 min read
Why we built Penwright local-first
Privacy, latency, and the engineering bet on whisper.cpp + llama.cpp + Metal. The opinionated essay version.
Engineering·7 min read
How Penwright cleans up your dictation
An honest tour of our three-tier on-device cleanup pipeline, rules engine, small local model, EuroLLM 1.7B (native EN + DE), and why we don't reach for a 7B+ model for every sentence.
In the queue
Why 150 WPM isn't the ceiling
What happens when typing disappears from the equation.
Drafting