madwatch

The statistical core of the anomaly detection engine, extracted into an open-source Python package. Median-based, so one whale transaction can’t inflate the baseline and mask the next real anomaly.

What’s inside

mad() and modified_zscore() as a pure-NumPy core, a RollingDetector for streaming data, and a SeasonalBaseline that buckets history by day-of-week and hour — the same combination that cut false positives by 60% in production. A CLI ships with it: madwatch data.csv --column amount --plot out.png.

How it was built

Test-driven from the first line: 39 tests including the whale test, CI on four Python versions, releases flow to PyPI through trusted publishing — no tokens, just a git tag.