Codeball

Thanks @rememberlenny and for sure, happy to share! Codeball is a Multi-layer Perceptron classifier neural network. It evaluates on a per PR file basis, taking in a very large number of inputs about the code and the author. Broadly speaking, they fall in three different categories — Basic, Derived and Categorical. - Basic — Metadata directly associated with the Pull Request. For example, number of lines added, removed etc. - Derived — Data that is indirectly associated with the Pull request. Examples include number of days since last change to file, file ownership ratios, file volatility score, author-file score etc. - Categorical variables — Derived categories for the Pull Request, for instance the file type categories. A more advanced example category is the perceptual hash of file diffs. We used millions of pull requests from open source projects in order to generate training data - the predictors as well as the known outcomes (whether the PR was merged without changes/objections). Having seen tons of success and failure patterns makes Codeball really precise in evaluating contributions it has never seen before. Right now Codeball can only classify (and trigger approval of) safe PRs. We want to take it further and give the developer feedback on how to improve their code. Here is a comparison with the automation level for autonomous cars ;) -> https://codeball.ai/what

Thanks @mariel_davis