# AuditGlimmer

- **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:** [Shroddha Ghosh](https://cerebralvalley.ai/u/shro-2002)
- **GitHub:** https://github.com/shro-2002/auditglimmer
- **Demo video:** https://youtu.be/xvGoop1LOUE?si=letmXDoImgIGVmZT
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/314

AuditGlimmer — Turning Inspection Photographs into Traceable Financial Evidence**

When financial institutions lend against physical equipment, verifying collateral is a critical but largely manual process. Inspectors collect photographs of machinery and nameplates, while auditors must determine which images support which assets in the lender's register. Similar-looking machines, unreadable serial numbers, and conflicting records make this process difficult and increase the risk of incorrect decisions.

AuditGlimmer is a multimodal AI-powered collateral evidence and reconciliation platform built using Meta SAM 3.1 and Muse Spark 1.3.

The system combines two complementary Meta AI capabilities:

Meta SAM 3.1 segments individual machines in site photographs, allowing evidence to be associated with specific equipment regions.
Meta Muse Spark 1.3 extracts visible manufacturer, model, and serial-number information from equipment imagery, while recording uncertainty when identifiers cannot be read reliably.

A deterministic reconciliation engine compares the extracted evidence against a lender's asset register, identifies potential matches and conflicts, and routes uncertain cases to human reviewers. Every decision is linked to its supporting photographs, model outputs, reconciliation rules, and an auditable record.

What makes AuditGlimmer different?

Instead of treating AI-generated answers as proof, AuditGlimmer separates model observations from financial decisions. A readable serial number can identify a candidate asset, but does not automatically establish ownership or physical identity.

In a genuine Meta API demonstration, Muse Spark extracted serial number `28756` from a real nameplate. The system narrowed a 44-row demonstration register to one candidate, correctly preserving the decision as **PROBABLE**, pending human review.

In a separate 14-photo holdout study, Muse extracted 5 of 6 legible serial numbers exactly. A retrospective comparison showed that incorporating Muse's readings improved correct single-candidate suggestions from 0/6 to 5/6 compared with rules alone. These results are experimental and do not establish production-level accuracy.

Our goal is to make physical collateral inspection more traceable, explainable, and efficient—while keeping humans responsible for consequential verification decisions.

## More from Meta Global AI Developer Hackathon

- [Rockey kotwal](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/311)
- [coco](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/312)
- [Aula Raíz](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/313)
- [SahiBill](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/315)
- [Beta Labs](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/316)
- [FORGE](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/317)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/314. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt