It uses a BERT-family model to cluster and classifies articles, it learned very fast that I like articles that have to with recommendation systems, text classifiers and stuff like that and I don't like articles on numerical analysis. The main interface looks like TikTok (shows me an article that I "like", "don't like" or mark as favorites.) Once things are favorited I have a rudimentary interface to look them up.
The same parts can be put together in other ways, the system right now supports one stream of classifications, but I'd really like to do a search on pubmed that finds 5000 abstracts and then classify them, plus multiple-stage classifications and extractions such as "Is this article about the results of a sports game?" and then "Who won?"