Music Man — Development skill for Claude Code

Music Man — Music-Streaming Royalty-Fraud Lakehouse

An agent that investigates suspicious music-streaming patterns (bot-farm royalty fraud) and proposes withholding a payout — a human always approves before anything actually happens. Built on PySpark + Delta Lake + MLflow, running entirely locally, portable to a real Databricks workspace.

Companion project to [NorthStar's Agentic Intervention Copilot](../northstar-platform): same human-in-the-loop spine (agent proposes, never executes), different stack — this one proves out lakehouse-scale data engineering instead of a React/Next.js frontend.

Results (a real run, not a mockup)

Verified against the actual public catalog and a full pipeline run:

• 89,740 real tracks, 31,428 real artists seeded from a public Kaggle Spotify dataset
• 2.1M synthetic play events processed through the full bronze → silver → gold medallion pipeline on local PySpark + Delta Lake
• 40 injected fraud cases across 4 distinct bot-farm signatures — the trained classifier caught 100% of them (recall 1.0) on a held-out test split, at 67% precision. The false positives are genuinely popular artists whose scale overlaps with the fraud signature — which is exactly the argument for why a human reviews every hold before it takes effect, not a flaw to hide.

**A real excerpt from a live agent run** — it didn't just call tools in sequence, it caught a problem in my own synthetic-data generator mid-investigation:

*"device_concentration_ratio is 0.0668 here and 0.0667–0.0669 for every other top candidate inspected (Cachureos, King 810, Feid). That near-identical value across unrelated artists looks like a pipeline artifact rather than an independent signal, so it should NOT be weighted as corroborating evidence."*

Unprompted, it also noticed the model's anomaly scores were saturated at 1.0 across 39 artists, refused to rank candidates by score alone, tie-broke on financial exposure instead, and flagged that the shared signa