Draft:Metabob

Metabob is a software platform for static code analysis. Metabob is an ensemble AI system for classifying, identifying, and explaining non-deterministic faults within source code. Metabob uses a combination of graph-attention neural networks (GNNs) and generative AI to improve software performance and security.

Technology
The software uses topic modeling to build seed data sets. Here, the underlying reasons behind particular classes of code changes are extracted from the surrounding documentation behind each code change. This allows to conduct supervised training of a classifier using an extended version of the Abstract Syntax Tree (AST). This is parsed from the source code and used as the input vectors to a GNN. The fault class, as determined by the topic model, is used as the output class in the GNN.

Metabob then generates explanations and code suggestions for fixes via a language model. These are built on a context vector from the topic labels, the source code, and portions of the online documentation, docstrings, headers, and other non-local information (readme’s, etc.). This results in simple explanations of the underlying issue behind a particular problem and related code change recommendations.

In consequence, Metabob can detect context-based software bugs that aren't identified by traditional, rule-based, static code analysis tools. Metabob examines entire codebases and identifies errors that are results of various code fragments. The detected problems range from race conditions and memory leaks to unhandled edge cases (among others). A recent study showed that Metabob's AI code review significantly reduces debugging and refactoring times.

History and Awards
In early 2023, Metabob launched its VSCode extension. The software is currently also available through GitHub. As of Fall 2023 the AI model supports Python, JavaScript, and TypeScript.

In 2023 Metabob was awarded an SBIR (Small Business Innovation Research) Phase I grant from the National Science Foundation and won the 2023 AI TechAward in the category Deep Learning Technology.