Machine learning / 2026 / Live in the browser
Halo
A reverse GitHub for AI-written work. Give it a repository, gist, pull request, package or web page and it returns a calibrated estimate of whether the work was AI-generated, with the file, commit or line that shows why.
01
The problem
AI detectors give one score with no reasons, and most only read GitHub, so a verdict cannot be checked and anything hosted elsewhere is invisible.
02
How I approached it
Repositories are read through the git protocol alone, so any host works and host APIs only add detail. Declared involvement such as Co-Authored-By trailers, agent bot accounts and CLAUDE.md files is reported apart from inference. 18 statistical detectors cover commit timing, history mess, message style, code stylometry on tree-sitter ASTs for 10 languages, duplication against a template corpus and drift from the author's own pre-2021 writing. Their outputs are fused as weighted log-odds learned by logistic regression, calibrated, and given a 90 percent bootstrap interval. Vercel's Python runtime has no git, so the build copies a git binary into the function.
03
The outcome
Two models: a repository model trained on 80 real repositories, and a prose model for web pages trained on 1,956 HC3 and MAGE texts across 16 domains and 35 generators. Every report lists its top evidence with file, commit and line pointers, and limitations written from whatever data was missing.
Related work
A neural network that learns chess by distilling Stockfish, with MCTS on top and a proper Elo measurement against Stockfish itself.
An image classification pipeline trained to recognise meme formats from a curated dataset.
A 26 KB neural network running entirely in a browser tab. Draw a digit, it classifies on device, and the page reports which execution provider it actually got and what the inference cost.