# NightMoose (solo: Alex Suarez, GitHub @nightmoose)

- **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:** [Alex Suarez](https://cerebralvalley.ai/u/nightmoose)
- **GitHub:** https://github.com/nightmoose/pipeline-medic
- **Demo video:** https://youtu.be/uGSQObJzdxs
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/239

Data pipelines break at night: a column gets renamed upstream, a type drifts, a join starts duplicating rows, or nulls flood in. An on-call engineer then spends the morning reading logs.

Pipeline Medic is an agent that handles that loop end to end:
1. It runs a real Airflow DAG over a real dbt project (dbt-core + dbt-duckdb) and catches the failed task.
2. It collects context from dbt's run_results.json and manifest.json: the failing node, its SQL, the failing test or contract, row and null stats, and the upstream columns with sample rows.
3. It asks Muse Spark 1.3 (Meta Model API) for a structured diagnosis: root cause, evidence, confidence, a unified-diff patch, and whether to fix or only suggest.
4. It verifies the patch in a sandbox copy: dbt compile, a full dbt build, row-count and null-rate bounds against a clean baseline, and an Airflow DAG re-run.
5. It drafts the pull request (branch, commit, PR body with evidence and the verification log). In the demo this step runs as a dry run.

There are four seeded failures in the demo repo. In the recorded live run (artifacts in demo_run/), Muse Spark diagnosed all 4. The 3 code fixes (upstream rename, schema drift, bad-join fan-out) passed sandbox verification on the first try. The null spike was correctly flagged as an upstream data issue, with no code patch.

Limits, stated plainly: this is a hackathon demo on a small seeded project, not a production deployment. PRs are drafted as dry runs, and the live results come from one recorded run.

## More from Meta Global AI Developer Hackathon

- [MARMAR Health](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/236)
- [FraudPause](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/237)
- [FaultForge](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/238)
- [Vestire AI Listing Studio](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/240)
- [Watch Me](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/241)
- [VERA](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/242)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/239. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt