# Swanix 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:** [Gurupriya Arindam](https://cerebralvalley.ai/u/Swanix)
- **GitHub:** https://github.com/Arindam5294/unravix-lens
- **Demo video:** https://youtube.com/shorts/WP1mjmTFl3k
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/403

Unravix Lens — Snap your working. Lens finds the exact line where it went wrong.

Students practise on paper. When an answer doesn't match the key, they usually redo the whole problem, which hides where their own reasoning broke, so the same slip comes back in the exam. Unravix Lens finds the knot: the first line where the student's thinking went off track.

- Check my working: photograph any handwritten step-by-step work, from Class 1 arithmetic to PhD proofs, physics, chemistry, statistics, accounting, code traces and grammar. Lens boxes the first wrong line on the photo and names the misconception in plain words at the student's level.
- Hints before the fix: three graded hints come first, and the corrected line stays hidden until the student asks. Fix it on paper and rescan to confirm the knot is untied.
- Carried errors are labelled, so the student fixes one thing, not five. Three tap-to-reveal check questions follow.
- Knot Report: a private history in the browser that surfaces weak spots that keep coming back and writes verified practice problems.
- Ask a doubt: type, photograph or speak any question and get the answer, short steps, the key idea, the trap to avoid and three check questions.
- 65 languages in the interface and answers, with formulas left as written. Phone-first: on-device auto-rotation, no install, no login.

Live demo: https://unravix-lens.pages.dev
Public benchmark: https://unravix-lens.pages.dev/benchmark.html
Built with: Meta Muse Spark 1.3, Meta SAM 3.1, Meta Muse Voice Transcribe, Cloudflare Pages, Cloudflare Workers, Workers KV, TypeScript, JavaScript.

ACCURACY:
Public held-out benchmark of 60 items (Class 3 to PhD, 11 subjects), written after tuning and never used for prompt work: 32 of 36 planted mistakes found, all on the exact line; 0 of 20 clean pages falsely flagged (1 of 20 before the second look); 4 of 4 essays correctly left unmarked; 56 of 60 overall, about 13.6 s and 2.4 cents per item. Every image, result and miss is public. On real notebook phone photos, several taken sideways, every visible real knot was caught on the exact line. Known limits: one missed dilution slip, two mistakes flagged one line late, and occasional strictness on notation.

TECHNICAL EXECUTION:
Cloudflare Pages, then a Pages Function and Service binding, then a TypeScript Worker calling the Meta Model API. A spend guard warns at $100 and hard-stops at $140. Per-device daily limits are stored as SHA-256 hashes, with a separate judges' allowance. Photos are resized and stripped of EXIF/GPS on the device, sent once and never stored. 146 unit tests, strict TypeScript.

IMPACT:
For any student who practises on paper, from Class 3 arithmetic to entrance and university exams, especially where teachers have little time for one-to-one checking. The student stops redoing whole problems and fixes the one idea that broke. It costs 2.6 to 5.6 cents per real-photo scan and about 1.2 cents per doubt, cheap enough for a school or coaching centre to offer free. The same page can be explained in Hindi, Tamil or Arabic.

CHALLENGES:
1. SAM returned no line boxes on most lined-notebook photos, so we kept it as an exact anchor and built an on-device ink snap.
2. Spark's boxes drifted down the page and hopped lines; an order-preserving alignment that scores consistent drift fixed it.
3. Sideways photos were slow and costly until we added free on-device rotation.
4. Crossed-out ink caused false flags; the second look fixed them without losing real catches.
5. With no provider spend cap, we built our own budget guard.

ACCOMPLISHMENTS:
The red box lands on the exact line on real, sideways, dim phone photos. 56/60 on a public held-out benchmark with zero false flags on clean work, misses included. A tutor that coaches (hints, fix, rescan) instead of giving answers away, in 65 languages and by voice. Built solo in a 10-day hackathon.

WHAT WE LEARNED:
Models are strongest when each does one job: SAM to see structure, Spark to reason line by line. Plain on-device geometry handles the last 5% of localisation. And honest evaluation beats a perfect-looking demo.

WHAT'S NEXT:
A bigger real-photo benchmark with schools and many students' handwriting; accepting common notation conventions without hiding real errors; a teacher view showing which misconceptions to reteach; voice answers read aloud in the student's language; folding Lens into the Unravix study app.

## More from Meta Global AI Developer Hackathon

- [Unamino](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/400)
- [Promptforge](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/401)
- [Connector Foundry](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/402)
- [fixalong](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/404)
- [Atik Apps](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/405)
- [Debug Dynasty](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/406)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/403. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt