Can't speak for this particular application but just off the top of my head you could use NLP to aggregate news based on more complex queries. Instead of saying "grab the most popular article" you could say something like "grab the most popular articles that talk about both Hillary Clinton in a positive tone and Bernie Sanders in a negative tone and is about the economy."
You should be able to build that topic on idina. If you start with a search for "clinton sanders economy," then communicate the "clinton good, sanders bad" by rating examples of that high, and everything else low.
It is hard for algorithms to understand words like "data-driven political journalism that is slightly left leaning," and, but easier to understand "stuff like this couple of Nate Silver article." So for the short term, I think we'll continue to need positive and negative examples to really get topics good for people, but someday I'd love to figure it all out from NLP parsing a general description.
For customized categories, the machine learning is optimizing for ordering the articles as closely as possible to how the user rated them. There are other options, like optimizing to clicks, but that can end up with more tabloidy articles than most users like. Our hope is by optimizing to explicit ratings, users will be able to find content that is more deeply interesting.