I publish code on GitHub as NetBr3ak.
I build offline speech software for Apple computers and run language-model experiments on one GPU. I founded Tenzan Logic, which makes Voxmill and Helio.
Software (closed source)
- Voxmill: offline transcription for Apple computers, built with Rust, whisper.cpp and Core ML. 7.74% word error rate on the 758 read-speech clips of the Polish FLEURS test set (Voxmill Pro, Whisper large-v3-turbo Q5_0, normalised text, measured by Tenzan Logic, September 2026).
- Helio: a demo that analyses solar farms from the inverter data they already log.
Open source
- BETON: a Claude Code plugin. After each
Write,EditorMultiEdit, it runs the installed linters and type checkers, and tests matched by file name. It also blocks shortcuts such as# noqaand skipped tests. - withastro/astro #18660, merged: preload variable fonts when the requested weight range is inside the file's range.
Research
HSPMN is Python research code for an unpublished manuscript on hybrid language models. It compares a learned per-token gate with a frozen random gate for routing between a Gated DeltaNet recurrent stream and a Native Sparse Attention stream. I trained models on WikiText-103 at two scales, 140M and 350M; these are scale names, not exact parameter counts (recorded findings).
A later audit found a bug that shifted the training targets twice, so every recorded run learned to predict the token two places ahead, not the next token. The audit also found future-token access in the predictive-coding gate and automatic NSA block selection, and an incorrect Gated DeltaNet CPU recurrence. Version 5.1.0 fixes these defects; I have not yet run it at training scale (reproduction status).
Manuscript source (LaTeX), written before the audit.