| name | drift-detect |
| description | Detect data drift, concept drift, and model performance degradation in production. Uses PSI, KS-test, and chi-squared for statistical drift, plus evidently and nannyml for automated reports. Use when monitoring a deployed model or comparing training vs production data distributions.
|
| allowed-tools | Bash(uv run * scripts/detect_drift.py *) Bash, Read, Write, Edit, Glob, Grep
|
| argument-hint | reference dataset and current dataset paths (e.g. "data/train.csv data/production.csv") |
| model | sonnet |
| effort | medium |
| compatibility | >=1.0 |
| metadata | {"category":"mlops","tags":["drift-detection","monitoring","psi","ks-test","evidently","nannyml","production"],"phase":"monitor"} |
Drift Detect Skill
Detect data drift, concept drift, and model degradation in production.
Quick start
uv run ${CLAUDE_SKILL_DIR}/scripts/detect_drift.py data/train.csv data/production.csv
Drift types
| Type | What it measures | Tool |
|---|
| Data drift | Input feature distribution shift | PSI, KS-test, chi-squared |
| Concept drift | P(y | x) relationship change |
| Target drift | Label distribution shift | evidently TargetDriftPreset |
| Model degradation | Performance drop in production | nannyml CBPE |
Thresholds
- PSI > 0.2: significant drift — investigate immediately
- PSI 0.1-0.2: moderate drift — monitor closely
- KS p-value < 0.05: statistically significant distribution shift
- Chi-squared p-value < 0.05: categorical feature drift
Integration
Add drift check to inference pipeline:
from scripts.detect_drift import compute_psi
psi = compute_psi(reference_col, current_col)
if psi > 0.2:
alert("Significant drift detected")
See references/drift-guide.md for complete PSI formula, evidently, nannyml, and alert patterns.
