Pipeline diagram showing the flow from PR metadata through style analysis to final review generation:

I built this to make code reviews more consistent without losing each reviewer’s personality. The pipeline pulls PR metadata, diffs, comments, and past reviews, runs a quick local pass with Qwen 2.5 (via Ollama) to filter out anything sensitive, and then has Claude 3.5 Sonnet extract the reviewer’s tone, recurring themes, and technical focus areas into a crisp set of instructions. Under the hood it’s orchestrated with Haystack so components can be swapped easily.

From there, I can feed a new PR’s metadata and diff through the notebooks to generate a draft review that sounds like the target reviewer while staying technically accurate. There’s a small helper script (fetch_pr_diffs.py) that works off a local clone to grab diffs efficiently. Notebooks cover three steps—classify, learn style, apply style—so it’s easy to audit each phase. Repo: pr-review-style-imitate.