AI Video Pipeline, API-only
FSL-1.1-MIT, source available
Turns any clip into a post-ready vertical video: a voice swap that keeps lip timing, word-timed captions, QA sheets and a Remotion end card.
Run it
npm test python pipeline/finish.py my-clip.mp4 --keep-audio

Word-timed captions, from the self-test clip
- Loudness
- -14.0 LUFS
- Target
- -14 LUFS
- Vertical file
- 1080 x 1920
- Rate
- 30 fps
finish.py exports; captions burned in. Word times from out/vo_work/_selftest.srt; size and loudness from npm test, 18 Nov 2026, re-run 20 Nov 2026What it does
The finishing half of a short-video factory. Bring a clip from any AI video tool or your phone; it exports 1080x1920 at 30 fps with metadata stripped and the sound normalised to -14 LUFS.
The full description
A claims lint checks captions and ad copy for health or guarantee claims, stale discounts, fake-testimonial phrasing and income claims.
Official APIs and local tools only: ElevenLabs, ffmpeg and Remotion, with no browser automation and no scraping of any video service. Every API result is cached by input hash, and a dry run prints a credit estimate first.
What it caught
From the offline test run and the claims lint, with no API key:
python pipeline/claims.py "Guaranteed results, 50% off"
FIX: health or guarantee claim; exact discount claim (goes stale); say 'on sale today' or show the live price
exit 1claims.py flags a health or guarantee claim and an exact discount that goes stale, and exits 1.
python pipeline/claims.py "Fresh bread every morning"
OK
exit 0The same lint passes plain copy and exits 0.
- No key, no callvo.py refuses every ElevenLabs command until your own key is set, and never prints a key.
- A QA sheet of every frame that mattersqa.py writes a timestamped contact sheet every 0.4 s, or every 0.1 s for the first 2 s in dense mode, to catch melting text and extra fingers.
Source: pipeline/claims.py, 20 Nov 2026; npm test (tests/run.py), 18 Nov 2026, re-run 20 Nov 2026; README.md for qa.py