To get feedback at scale you could, on the relay side, make predictions of what Engagement is expected on each note (like, zap, comment) and tweak the algo based on what actually happens on the note. Compare actual vs predicted and rerun your recommendations. Is that what you are doing? 

 No that would take way too much compute for every single user I think, more than my $6 VM can handle lol 

 Don’t you have to be making some kind of prediction of what you think the user will want inorder to send user specific content? They liked this before I think they will like this thing which is similar to it? 

 Seems pretty good, so far I will keep using it.

Nostur does use the “since” filter on relay feeds, and it won’t add older posts above newer posts when updating the feed, not sure if that breaks anything 

 It doesn't break it, the relay disregards since/until filters, but it does provide a specific order that's not timestamp based (score based) - so it should work much better if you trust the order the relay gives you