We're in the process of preparing to launch the automatic labeler feature - things are looking good and we just need to make a few tweaks to have a working beta available (we'll also be reworking the site so things will look a bit different).
And quickly touching on the ML (great point, by the way): for the labeler, we only process each individual issue as it's raised so it would be independent of prior activity taken in the repository issues. The contributor assignment does look at history (since we try to get an idea of who can best handle what); we're evaluating some changes to this feature to make it more flexible so we'd definitely love input on that as well.
We'll reach out once we have the labeler ready for use. Thanks so much you guys!
P.S. To answer your question from the GitHub issue comment chain, Victor: currently we do not utilize cross-training in our models but since the labeler takes each issue as an independent event, you can use as many repositories as you feel are necessary with Heupr.