Comment by harlequinetcie

The strategy is quite straight forward:

-> Clairic extracts the PR's claims -> Jev tags diff hunks, then judges what requires a review. -> Cheap(er) model to review the hunks.

For us, we've got PR reviews in under 30 seconds, and cost down to a few cents.

Next step, we are thinking on training some small open weight models specialized on our repo.

Thoughts?