# Rewind Labs

- **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:** [JEFFERSON SANTOS AREQUE CHIRINOS](https://cerebralvalley.ai/u/devps)
- **GitHub:** https://github.com/JeffersonAreque/rewind
- **Demo video:** https://youtu.be/d6TLUnhGgak
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4128

Thousands of businesses still run legacy software with no source code, no documentation and nobody left who understands it. Modernizing it usually starts with months of reverse engineering.

Rewind starts from what every business still has: someone who knows how to use the old program. You record a screen video of a normal session, and four agents turn it into a modern, verified web app:

1. Screens: reads every window, field, button and table from keyframes.
2. Workflows: reconstructs what the user did step by step and extracts business rules (required fields, error codes, state changes, calculations), each backed by evidence steps from the video.
3. Builder: generates a modern web app from a validated JSON spec (the model never writes free-form code), keeps every legacy field and rule, and proposes modernization changes justified by the recording.
4. Verifier: replays every recorded workflow on the new app with Playwright and compares each step with the original keyframe using vision. Unclear is a warning, never a pass.

One command (rewind run video.mp4) produces the app and a single HTML report: before/after screens, workflows, rules, changes and step-by-step verification.

Results on two different legacy apps:
- Billing system: 29/29 steps, 10/10 rules, 4/4 workflows replayed, 25 pass / 0 fail / 1 unclear.
- Clinic appointments, recorded at real user speed (fast typing, a typo, a forgotten field): 41/41 steps, 4/4 workflows, 36 pass / 0 fail / 2 unclear.
- Deliberately broken copies of the apps (no balance formula, no required field, wrong default) were caught 3 out of 3 times on both apps. One break was caught only by the vision comparison.
- 136 tests, and re-running a demo costs 0 tokens thanks to caching.

## More from Meta Global AI Developer Hackathon

- [Padhai Saathi](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4125)
- [Runtime Terror](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4126)
- [ChangeProof AI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4127)
- [Team ID: 656295919540867](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4129)
- [Ayai](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4130)
- [ByteToBit](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4131)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4128. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt