02 · Brand film
58.2 s · 1280×720 master · 1080×1920 vertical · 30 s cut · live on the product site
A landing-page film for an AI companion worn by elderly users and monitored by their families. Two audiences, one runtime, and a subject where anything that sounds synthetic or salesy destroys the trust the product depends on — with no budget for a shoot, and a duty of care that rules out filming real families.
Fully AI-generated performers and scenes, composited against exact real screen captures, with narration that is not prompted text-to-speech. The founder read the locked script three times; the strongest read became a timing track, and a licensed narrator voice was conformed to that human cadence phrase by phrase. Pauses were then recomposed against picture so the film opens spacious and tightens as it closes.
Three deliverables from one build: a 58.2-second 16:9 master, a 1080×1920 vertical, and a 30-second cut assembled from the same verified audio — word-identical to the long version, not a re-record. Narration passed a 90-of-90 word-level accuracy check before picture lock. Live on the product's landing page.
The narration is five contiguous slices of the human performance. None is time-stretched, so the cadence inside each phrase is untouched. What changed is the rhythm between them:
| Gap | Human read | In the film |
|---|---|---|
| 1 → 2 | 1.25 s | 3.65 s |
| 2 → 3 | 0.54 s | 6.20 s |
| 3 → 4 | 1.25 s | 0.29 s |
| 4 → 5 | 1.38 s | 0.25 s |
Those widened early gaps are scored, not empty — the four character-dialogue spots land inside them, so the narrator steps back to let a scene play. The two voices sit 6 dB apart throughout, because the product is an AI that whispers to the wearer and the level differential is how that reads on the ear. Music is sidechain-ducked 8:1 under dialogue.
The word-accuracy gate. Narration was checked against the locked script by automated word-level speech-to-text before picture lock. The first voice render scored 78 of 90 words. The twelve missing words were repaired from an approved earlier take — rather than editing the script to match the render, which is the easy path and the wrong one.
Seven versioned builds, 3,219 lines of deterministic build code. Every frame and every word is reproducible from source.
Script lock · founder voice direction and recording · cadence mapping · licensed narrator conform · exact-word QA gate · screen compositing · caption system · music mix and mastering · 16:9 / 9:16 / 30 s deliverables · reproducible build.
Next → BarCraft AI — Party Panic