Spam filter for a hobby forum

Design spam filtering for an online forum for amateur astronomers with 400,000 members and about 25,000 new posts and replies a day. Spammers post links to scam shops, often from new accounts, and every few weeks a new wave arrives from fresh domains. A trained text classifier is available, and 12 volunteer moderators can review a limited queue each day. A clear spam post should never appear publicly, a legitimate post should not be delayed by more than a second or two, and moderator decisions should improve the filter over time. Describe where the model runs in the posting flow, how borderline posts are handled, and how labels flow back into training.