1) tldr summary of each post (when a link is first submitted). 2) Deduping posts. 3) Detecting when there are multiple stories (different links) that are basically covering the same thing, in which case it'll "merge" them by using only the best post (it decides). 4) Come up with titles for posts that don't have them (like twitter/bluesky posts).
Jev is used for:
1) Quickly getting rid of bad posts during the 1st stage of crawling. 2) Determining initial amount of points (based on a scale) awarded to post when it was first submitted.
There are probably few other cases I'm not thinking of right now. But basically, the gist is to use Jev for super frequent classifier tasks that are early in the pipeline so it can save cost.
This will roughly use about $40 of my AI credits per month. Anthropic gives Ultra users $200 monthly credit for API usage. So I'm happy I can make some good use out of it.