# DataTraces

- **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:** [Musharraf Shaikh](https://cerebralvalley.ai/u/mmshaikh03)
- **GitHub:** https://github.com/musharraf3/DataTrace
- **Demo video:** https://youtu.be/TsXX8NXRmlw
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4089

Every month, healthcare finance, actuarial and analytics teams get the same question before the cost review: paid claims jumped, so is it real? Claims data makes it hard. Late claims keep arriving for weeks (run-out), adjudication is pending, reversals net against paid, and eligibility files get rebuilt. A genuine trend and a data defect often land in the same month.

DataTrace gives that question to a Muse Spark agent with real tools. Given the previous and current SQL, two data snapshots and the metric definition (or just an actuary's voice note and a dashboard screenshot), it:

- replays old and new SQL on old and new data;
- tests hypotheses (duplicate eligibility rows, members dropped by joins, service-date boundaries, pending claims, reversals) with bounded, read-only SQL in a local sandbox, never seeing raw rows;
- reports a fix only after a deterministic checker verifies the complete per-claim result;
- gets an independent review from Muse Wonder, which can send it back to revise;
- pauses and asks when the definition is incomplete, instead of guessing.

On real-format CMS DE-SynPUF Medicare claims, it showed that of a +$5.21M jump in November paid claims, $4.56M was genuine run-out and $649,400 was an eligibility-rebuild defect across 69 claims. It produced verified corrected SQL and a leadership brief. The same engine runs retail revenue incidents through a domain profile.

Across 360 paid benchmark runs against the same model given every raw row and no tools, the agent scored 179/180 versus 168/180. More importantly, it put the right dollar amounts on real change and reporting error in 16/16 claims runs versus 4/16. At 42k rows the no-tools approach cannot run at all (context_length_exceeded). In a planted-error test, the reviewer caught 28/28 errors.

## More from Meta Global AI Developer Hackathon

- [SebastianMejia](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4086)
- [X-Arc](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4087)
- [Josue Arcos](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4088)
- [Abel Mendoza](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4090)
- [Praxis](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4091)
- [EvalAI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4092)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4089. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt