AI video studio for the opening seconds of an ad
A platform that generates video hooks with the advertiser's own face and voice, orchestrating text, image, video and voice models with an approval at every step.
- Industry
- Advertising
- Engagement
- My own product, prototype stage
- My role
- Product, architecture and development
- Stack
- Python, FastAPI, Dramatiq, PostgreSQL, pgvector, Next.js, Claude, ElevenLabs, Kling
The problem
The first seconds of an ad decide whether people keep watching. Shooting lots of variants with the person who fronts the brand takes time and money, and anyone running campaigns needs to test many.
What I decided
- Approval between steps. Video is the expensive part, so nothing is generated until the script and the image are approved.
- A visual judge built on Claude that checks every image before any money is spent on video.
- Consent as part of the product. The avatar can be revoked, consent records live in storage that can’t be deleted, and every video carries an AI-generated content label.
What I built
- A “director” with structured output that turns a brief into a scene and a voice-over script, and refuses requests involving third-party brands or minors.
- A PostgreSQL state machine that orchestrates the image, video and voice models and survives restarts mid-process.
- Face likeness checks, ffmpeg editing normalised to −14 LUFS, and subtitles.
- Cost tracking for every approved hook.
Result
Prototype deployed since September 2026. Voice-over hooks are validated; talking-to-camera hooks are in testing. Each approved hook costs $4 to $8 in model usage.