# Synthvale

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Leo Hu](https://cerebralvalley.ai/u/chula)
- **GitHub:** https://github.com/Hothit1/musecut-Agentic-video-editor
- **Demo video:** https://youtu.be/wwhqyyPrXg0
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/132

MuseCut is an agentic video editor that turns a long talking-head video into a ready-to-post 9:16 short, with the creator in control at every step.

The problem: creators spend hours turning one long video into a short. They rewatch the footage, pick the strongest moments, cut filler and pauses, reframe to vertical without cropping the speaker's face, add captions, find B-roll, design a thumbnail and write the post. Fully automatic clip tools are fast, but they give a black-box result the creator can't fix before publishing.

How MuseCut works: upload a 5–20 minute video and describe the short you want (length 30, 60 or 90 s, tone, key points). MuseCut drafts an Edit Plan:
• Story structure: a hook for the first 3 seconds, then the main points, then a closing that ends on a complete sentence. Every segment has a role, a one-sentence reason and a 1–5 score.
• Filler words removed from captions and long pauses trimmed, without speeding up the voice.
• A face-tracked 9:16 reframe that always keeps the speaker's face inside the frame.
• B-roll from the source video itself (slides, close-ups) or illustrative stills from Muse Image with a slow Ken Burns move.
• Word-highlight captions placed under the chin, royalty-free music that ducks under the voice automatically.
• 3 thumbnail options, plus title, on-screen hook, post caption, hashtags and chapters for the source video.

Then the creator edits the plan: keep, drop or reorder segments, fix caption text (timestamps stay locked), swap or remove B-roll, pick a thumbnail, toggle music, switch between 30, 60 and 90 s (the plan re-trims the weakest segments). Only after Approve does the plan lock and render. Nothing is rendered from an unapproved plan.

Output: a 1080×1920 MP4, an SRT file, the selected thumbnail PNG, the plan JSON, post copy, and a cost and time log showing which Meta model ran at each step.

Built for production: the plan is validated against a JSON Schema before it is saved, invalid model output goes back to the model with the errors (max 3 tries) and then fails with a clear message instead of guessing; render runs as a background job with queued / processing / done / failed states; network and rate-limit errors are retried with backoff; API keys are read only from the environment.

Impact: on a real 15-minute creator video (How To Make Everything, CC BY 3.0), MuseCut produced a ready-to-post short in 3.4 minutes including plan review, against the creator's estimate of 120 minutes by hand: about 35× faster, for roughly $0.21 in model calls.

Roadmap: dubbing, lip-sync, and multiple shorts from one long video.

## More from Meta Global AI Developer Hackathon

- [Sabi](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/129)
- [FixLens](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/130)
- [Moshimo Town](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/131)
- [HeyWalk](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/133)
- [rajameta](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/134)
- [Baby Bust](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/135)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/132. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt