# Zenico Technologies

- **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:** [Diego Lobo](https://cerebralvalley.ai/u/diegolobo)
- **GitHub:** https://github.com/DiegoLLBB/kernel-credit-lab
- **Demo video:** https://youtu.be/eDjBFLvNEFA
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2694

Kernel Credit Lab is credit monitoring you can audit.

Credit analysts covering investment-grade companies have to catch every change to a company’s debt: a new bond, a term loan disclosed in a 10-Q, an early redemption, a tender. These changes are scattered across thousands of filings. Miss one, like Amazon’s $17.5bn term loan due 2029, and your picture of what the company must repay each year is off by billions.

Kernel reads every new SEC filing for 16 US investment-grade companies twice a day. A Muse Spark agent works through each filing with tools that enforce the rules: every fact must be an exact quote found word for word in the filing, every change to the debt records must cite those quotes, and materiality is decided in code. Before anything is published, Muse Spark Open (open weights, on Baseten) checks the alert’s claims against the evidence and refuses what isn’t supported. Muse Wonder writes a daily brief from computed figures only, and SAM 3.1 locates bars & charts in investor-deck charts so Kernel can verify the values Spark reads. No model computes a number.

Every number Kernel publishes is backed by the filing, zero unsupported claims. On a replay of 29 real 2026 events, scored against answers the agent never saw, 12 of 14 material debt changes reached an analyst: 9 as published alerts and 3 escalated because the agent could not confirm them, rather than guessing. Only 2 were missed outright, both Intel judgement calls (a refinancing it judged net-neutral, and a stock offering). On the strict measure, the year and amount right, Kernel scored 7 of 14, more than twice a one-shot model given the same filings (3 of 14), which also keeps no records to audit.

Then we froze the agent on 2 Nov and let it run on new filings. Its first live alert was exactly that Amazon loan, disclosed in a 10-Q, which we had missed ourselves when building the test set. Every alert links to the quoted sentence, the recomputed maturity wall and a full trace of the agent’s steps, refusals included.

Live: https://kernel-credit-lab.web.app

## More from Meta Global AI Developer Hackathon

- [zaxyar](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2691)
- [PodAI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2692)
- [PRLens](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2693)
- [Code Ranger](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2695)
- [BaitWatch](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2696)
- [Irzelindo Joaquim Artur Salvador](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2697)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2694. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt