Most AI code review platforms are off-device, run on PRs and integrate with GitHub (e.g. Coderabbit). With Coducky I'm trying to do the opposite and focus on fast, cheap, on-device iteration. Part of what I am thinking about now is how to leverage on-device models as part of the local code analysis layer. This reduces costs for the end user and can also enable precomputation while idle.

I'm looking at integrating Apple Intelligence alongside deterministic analysis using tree-sitter and graph/vector retrieval. Today I'm working on a spike to de-risk some of the ideas I am thinking about. I'm going to first try to use Apple's Foundational Models to generate branch descriptions automatically in the background based on branch/ref labels, commit subjects, diff stats, changed paths, etc.