Content moderation

How this question breaks down

1. 1PolicyCore21 minA moderation model can only learn what the policy says, so the first design job is to turn a written rulebook into categories, severity levels and allowed neighbours that labelers can apply the same way twice, and to measure success as harmful views, not accuracy.
2. 2ClassifiersCore18 minA post is an image, a caption, the words drawn inside the image and sometimes a soundtrack, and the harm is often only in the combination, so Kiteline scores the whole post with a cascade. Perceptual hash matches against known media go first, cheap per-modality models run on everything, and a fused model runs where the cheap ones disagree or the post is about to spread.
3. 3ReviewCore21 minKiteline gets more items needing a human than its reviewers can see, so the queue is itself a model: it orders work by expected harm prevented per reviewer-second, routes each item to someone who can judge it, protects the people doing the judging, and feeds their decisions back without poisoning the training data.

Work through the parts in order and you’ve covered the full design. Each part is one topic.