Adult content platform

The challenge

Every user and every character must be an adult, and generated video fails in ways people miss at volume.

Adult platforms carry a higher burden of proof. Every user must be an adult, every character must clearly be an adult, and no real person's face can appear. Generated video also fails in ways a human misses at volume: a face that drifts between shots, extra limbs, or an extra person who should not be in the frame.

What we built

Age assurance, an automated five-check gate on every generated shot, and a spending brake, on serverless AWS.

Three layers. First, age assurance: a self-attestation step, server-side refusals for accounts that need review, and a regional third-party age check where the law requires it. Second, an automated gate on every generated shot that scores identity, continuity, anatomy, framing and prompt match, with face-embedding similarity for identity and a vision model for the rest; each shot passes, retries or goes to a person. A policy filter classifies explicit content against the platform’s rules. Third, a spending brake that refuses any paid generation call when the budget is low.

How we delivered it

We measured before we gated. Test runs showed that identity held at 0.82–0.84 similarity with one character but fell to 0.38 with two, and that 23% of frames had an extra face. Those results set the pass and retry thresholds and changed how scenes are written. The web platform runs serverless on AWS (CloudFront, Lambda, S3, SQS and DynamoDB), deployed with OpenTofu.