# Blistrue

- **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:** [Sai Naveena Sri S](https://cerebralvalley.ai/u/sainaveenasri)
- **GitHub:** https://github.com/Naveenasri02/describe-to-inspect
- **Demo video:** https://www.youtube.com/watch?v=l2UNVx0E-kA
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2778

Blistrue turns a webcam and a laptop into a QA station for tablet blister strips. QA describes a good strip in one sentence; the station checks every pocket, names the rule that failed, signals the machine, holds the line when something is critical, and drafts the batch record for QA to sign.

The problem
Blister packs fail in small ways that matter: an empty pocket, or a tablet in the wrong place. Both have caused real recalls (a 2019 US recall of packs that could contain an empty pocket; a 2012 recall of about a million packs). In India, the revised Schedule M GMP rules now reach small drug makers too, yet many still inspect strips by eye and keep records by hand.

How it works
- Teach in one sentence: Muse Voice Transcribe hears the rule; Muse Spark 1.3 turns it into numbered rules that QA approves. Five photos of a good strip teach the product.
- Inspect: SAM 3.1 finds the strip, every pocket and every tablet. DINOv2, frozen and running locally, judges each pocket against the taught strip. Plain rules in code make the pass/fail decision; no language model does.
- Hold the line: a critical failure (e.g. a different tablet) holds the whole line. Reject, hold and lock go out over Modbus TCP.
- Investigate: the Shift Supervisor agent on Muse Spark Open (Baseten) picks its own measuring tools and drafts a finding. It can only request a restart; QA approves with a password.
- Prove itself: before each batch, a self-test paints synthetic defects onto a good strip and must catch and name each one, plus a physical challenge strip. Until then, the station stays locked.
- Fail safe: if the network drops, every strip is held for a manual check.
- Record: QA signs the drafted batch record; every step is kept in a hash-chained, tamper-evident audit trail.

Results from the recorded bench run (3 Nov 2026, real strips, the run in the video)
- Self-test: good strip passed; 4 of 4 painted defects found and named.
- Challenge strip (different tablet in a corner pocket): rejected, Rule 5.
- Real strips: good passed; broken rejected (Rule 3); empty pocket rejected (Rule 2); half tablet rejected (Rule 3), and three rejects in a row held the line for the agent; different tablet held as critical (Rule 5); another product's foil, past its expiry, held at the print check.
- Voice command "inspect": the good strip passed.
- Audit trail: 36 entries, chain intact.
- Known limits: strips must lie the same way round as when taught; a half tablet is named "broken". This is a bench station, not a line-speed system.

Built during the hackathon, 26 Oct to 4 Nov 2026: the station, its model pipeline and the agent. Before the event: the concept and an early private prototype only; no earlier code is used. The video is real footage of the station on a bench with real strips, not sped up; cuts skip waiting time. The log line under the screen is the station's own model-call log from that run (201 calls, none failed). Narration: recorded human voice.

## More from Meta Global AI Developer Hackathon

- [Cinemind](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2775)
- [AuraX](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2776)
- [HeroGen](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2777)
- [the-sp](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2779)
- [Raiiotronix Block Studio + AI Tutor](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2780)
- [SpeakEasy](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2781)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2778. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt