# Glint

- **Event:** [Gemini 4 NYC Hackathon](https://cerebralvalley.ai/e/gemini-4-nyc)
- **When:** Sat, Nov 21 at 9:00 AM – 10:00 PM (EST)
- **Where:** New York, NY
- **Team:** [Giridhar Mula](https://cerebralvalley.ai/u/giddy_uppp), [Nihantth Naidu Kalisetti](https://cerebralvalley.ai/u/nihanthnaidu007)
- **GitHub:** https://github.com/nihantthnaidu007/glint.git
- **Demo video:** https://youtu.be/zpjzWQpxDj8
- **Gallery:** https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/46

Glint turns a screen recording of a bug into a fix it can prove, or honestly says it could not.

Real bug reports arrive as screen recordings and voice notes, but coding agents are built for text tickets. Glint takes only the video and the app's source. It (1) watches the recording and works out the steps, expected behavior and observed behavior, (2) writes a Playwright test and keeps it only if it fails 3 of 3 times on the buggy app, (3) writes 4 candidate fixes in parallel, (4) checks each fix: the test must pass, random interaction must cause no errors, the change must be small, and a calibrated visual check compares before and after screenshots, and (5) returns a verdict with evidence: "Verified", "Likely fix", or "No verified fix" plus the failing test.

Proof: we generated 28 single-line bugs in TodoMVC with a seeded mutator (none hand-written), excluded 3 before scoring, tuned on 7, and froze 18 test bugs, thresholds and 6 predictions before the scored run. A hidden differential fuzz oracle (200 random interaction sequences against the original app, 37 of 37 controls passing) decides what a true fix is.

Results on 21 test runs: a single Gemini 4 Neon call truly fixed 95.2% and shipped 1 wrong fix. Glint v1 fixed 66.7% and was right on all 14 fixes it claimed. After seeing that its visual check rejected correct fixes, we added an agreement rule (v2): 85.7% fixed, 18 of 19 claims right, 1 wrong fix. So the single call fixes more simple bugs. Glint's value is telling you which fixes are proven and giving you a failing test when they are not. 3 of our 6 pre-registered predictions missed. All of this, including per-run records, is in the repo.

Everything was built on Nov 21, 2026. TodoMVC is third-party MIT code used only as a test subject.

## More from Gemini 4 NYC Hackathon

- [@labs](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/43)
- [Dunkoma](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/44)
- [INTERLOCK](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/45)
- [Mozart](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/47)
- [Overheard](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/48)
- [Let There Be Sims](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/49)

---

Markdown version of https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/46. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt
