AI product Open source
A small local system-one model and code-readability tool that flags structural smells such as deep nesting, noisy comments, swallowed errors, poor naming, dense lines, special-casing, duplicated blocks, and cramped layouts. Its 1.58-bit weights are deterministically compiled without training data or gradient descent and run through pure NumPy or browser Float32Array workers, without regexes, tokenizers, or ASTs.
The model encodes source files as a full-file byte grid, applies quote/comment and delimiter-state masking, multi-scale ternary convolutions, identifier stem sketches, and two banded relative-position attention passes before producing document scores, per-line squiggles, hygiene dimensions, and code-archetype probabilities. It provides Python APIs, a CLI for paths, globs, directories, or stdin, JSON output, an RL reward for test-verified optimization loops, and an agent skill for tools including Claude Code, Codex, Cursor, Antigravity, and Copilot. Functional bugs, type errors, security vulnerabilities, and broad legacy-repository sweeps are outside its stated scope.
1 use taken from transcripts — each links to the moment in the video.
A small local model that flags readability problems in source code, including deep nesting, swallowed errors, and duplicated blocks; its weights are compiled deterministically without training data and it runs via numpy, browser workers, CLI, Python, or an agent skill.
1 in the library.