# ShortForge

- **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:** [Kamlesh Thakur](https://cerebralvalley.ai/u/kamleshthakur)
- **GitHub:** https://github.com/Kamu2000/shortforge-meta-hackathon
- **Demo video:** https://www.youtube.com/watch?v=1HRk4joXOSs
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3080

AI Spatial Director, by Team ShortForge (Framiq Technologies), Startups track, built for Build for Your World.

The problem: the best short-form material on YouTube is buried in long single-speaker videos — podcasts, interviews, talks, lessons. Turning one into a Short that holds attention is skilled manual work: find the moment, cut it, frame the speaker, write a hook that lands in the first seconds, caption it without covering the face, show what is being described, keep the sound clean. Generic auto-clippers centre-crop and hope: hooks get buried, captions sit on faces, broken renders ship silently — nothing checks the finished file.

What we built: give the Director one long talking-head video and it returns up to three finished vertical Shorts (1080x1920), each with a measured PASS/FAIL receipt for the actual rendered file. It transcribes the source, scores the transcript for the strongest moments, cuts the speaker out of every frame, and composes each Short as a layered stage, not a crop: treated backdrop, hook text BEHIND the speaker (the speaker occludes part of the lettering — that is what makes the depth read), speaker in front, captions always in front and never over the face. When the speaker describes something, a graphics layer shows it — a generated still, a context card, a backdrop scene — at the moment it is said, under a placement law that never touches the face. A post-render verifier then measures the rendered file against nine frozen checks; on a FAIL the system may attempt exactly one repair before re-measuring. A Short that still fails is reported as a FAIL — we never manufacture a pass.

The evidence ships in the repository. The final batch produced 10 receipted Shorts: 3 PASSED_AFTER_REPAIR, 7 FAIL — failures published alongside the passes. The verifier's proof pack agrees on 44 of 44 labelled cases, the test suite passes 626 of 626 tests, and anyone can re-verify the bundled showcase receipt offline, keyless, with the checker shipped in the repo. Measured cost, API calls only, across complete passing runs: $0.5070–$0.7223 per run (one long video in, three Shorts out), $0.169–$0.241 per finished Short. The entry was built entirely on the $150 event grant — $21.9867 across 850 ledgered calls, under a $145 hard cap enforced in code, zero personal funds spent.

Who it is for: single-speaker podcasters and talking-head educators on YouTube, 20,000–500,000 subscribers, posting 2+ long videos a week, already paying for an OpusClip-class clipper. Price: 30 credits per run — $2.90 standard, $1.90 founding — against that measured $0.51–$0.72 cost. A creator with 278,000 subscribers is open to testing it.

Only work built during the event (October 26 – November 4, 2026) is in this repository. ShortForge, our existing AI video studio, is pre-existing; its codebase is not in the repo and no pre-event code or assets were copied in. The Director is designed to become a ShortForge feature the week after judging.

## More from Meta Global AI Developer Hackathon

- [AXIOMH](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3077)
- [Ayé](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3078)
- [Osee11](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3079)
- [Lannister](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3081)
- [MAFA](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3082)
- [DuckyHelper](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3083)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3080. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt