Traditional content moderation relies on classifiers trained for a single modality; a distributed architecture increases invocation costs and makes results difficult to unify.