# Intenzo

- **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:** [Nilesh Gupta](https://cerebralvalley.ai/u/0xnilesh), [Sumith Prabhu](https://cerebralvalley.ai/u/sumith)
- **GitHub:** https://github.com/0xNilesh/intenzo
- **Demo video:** https://youtu.be/W393A5ZtUVE
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/705

Intenzo is the one-stop way to test a pull request's intent: does it actually do what the ticket asked?

THE PROBLEM
AI made writing code fast: agents now start over 40% of pull requests on GitHub (Octoverse 2026). But shipping didn't speed up. Teams with heavy AI use merge 98% more PRs and spend 91% more time reviewing them (Faros AI). AI code is usually "almost right": it compiles, its own tests pass, the diff reads well, but the cart total only updates after a reload. CI checks the code and reviewers read the diff; nobody checks the product against the ticket without pulling the branch and clicking through it by hand. The bottleneck moved from writing code to verifying it.

WHAT INTENZO DOES
Intenzo is a GitHub App. On every pull request that links a ticket, it:
- Reads the intent: the ticket's text and even the PM's narrated screen recording become acceptance criteria (Voice Transcribe + Spark 1.3).
- Finds each criterion's path through the app (Wonder) and records it in a real Chromium on the PR's preview and on main, each time as a brand-new user, with exact readings before and after the action.
- Judges every recording with two models, Spark 1.3 and open-weights Spark Open on Baseten. A verdict stands only when both agree and cite the readings; otherwise it's inconclusive, never guessed. SAM 3.1 outlines visual evidence, like a chart slice, on every frame.
- Posts the verdict on the PR with the video as proof, flags changes no one asked for, and with one click sends a failure to a coding agent (Muse Code) that pushes a fix, which Intenzo re-checks before the PM approves.

DOES THE EVIDENCE MATTER?
On IntentBench, our held-out benchmark (8 cases, 28 criteria), judging from the ticket and diff alone passed 33% of failing criteria. Intenzo's video plus two-judge setup passed 0%, and no arm failed a criterion that holds.

BUILT WITH CRAFT
Next.js Cloud on Vercel (Postgres, Blob, an outbox so every GitHub write happens exactly once) and a worker on Modal (Playwright, Chromium, ffmpeg). Six Meta models, each with a distinct job. A full run costs $0.07-0.21 and takes 1-3 minutes. Setup is a GitHub App install, a short config file and a ~15-line session hook.

Today it works for web apps on GitHub with preview deployments (Vercel). Developers stop clicking through branches, reviewers get evidence instead of diffs, and teams get an approval gate tied to the ticket plus a video record of every change.

## More from Meta Global AI Developer Hackathon

- [QuantProof](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/702)
- [DriveMatch](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/703)
- [Theotech](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/704)
- [Horia BEschea](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/706)
- [NDxGenius](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/707)
- [ChartCortex](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/708)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/705. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt