Phantom

vreview is code review for dubbed and AI-generated video. AI dubbing and video generation now produce more footage than people can check, and whoever reviews a dub usually doesn't speak both languages. Teams either ship unchecked work or pay for slow manual QA. You upload the original and the dub (or call the REST API, the MCP server or a webhook) and get back a review like a pull-request review: a verdict (request changes / needs human / approve), a score, and timestamped threads grouped as blocker, should-fix, needs-human and nit. Each thread has evidence and a fix instruction that a dubbing or generation system can act on, for example "delay this line by 0.86 s" or "restore 'Hoy' at the start of the line". The review page plays the original and the dub side by side with an issue timeline. Reviewers mark each finding accept or false positive, and that feedback is stored as future evaluation data. A live pipeline view shows every stage, sub-run, model used and cache hit. Under the hood, deterministic signal processing (speech detection, Whisper word timestamps, line alignment, timing and sync checks, mouth-activity analysis) locates candidate problems. Meta's Muse Spark then watches and listens to short clips of the original and the dub around each one, confirms or disputes it, and adds what the checks can't see, such as lip sync, voice fit and lost emotion. On our live 60 s Hindi→Spanish review it confirmed 4 issues, disputed 8 and found 1 new lip-sync problem.