Just to be clear, I think this part is a good idea - my suggestion was to go a step further and summarize the contents of those posts pulled out by your own model using AI, rather than just outputting the highlighted posts as-is one after the other.
I agree that just feeding an entire topic consisting of 50% junk into an AI summary probably wouldn't give great results. But if you use your signals to pull out the noteworthy posts and then summarize those, I think the result would be likely be pretty good. AI is really good at summarizing text, after all.