Upload your reference dataset and production sample. Get a drift score per feature and an immediate alert on which columns to investigate.
A model that was 92% accurate last month may be 74% accurate today. Know before your users do.
Population Stability Index computed per feature. Industry-standard metric trusted by financial risk teams and ML operations engineers alike.
Works on continuous numeric columns (age, revenue, score) and categorical columns (country, plan type, device) — all in the same request.
Columns with PSI > 0.1 (moderate drift) or PSI > 0.25 (severe drift) are flagged automatically. No manual threshold tuning required.
Compare your training baseline against any production snapshot — weekly, daily, or on every deploy. Just swap the CSVs.
Add a drift check to your model deployment pipeline. If PSI exceeds your threshold, fail the build and block the deployment automatically.
Drift is computed purely on input data distributions — no model weights, no prediction logs, no infrastructure access needed.
Three steps from datasets to drift report
Your reference dataset is typically your training data or a recent healthy production snapshot. The current dataset is your latest production sample.
For each column you specify, we compute the Population Stability Index by bucketing values and comparing frequency distributions between the two datasets.
Receive a JSON report with per-feature PSI scores, a drift classification (stable / moderate / severe), and a list of columns that require immediate investigation.
Start free. Scale when you need to.
PSI measures how much a feature's distribution has shifted between two datasets. It was developed by the credit risk industry and is widely adopted in ML monitoring. A PSI of 0 means no change; above 0.25 indicates severe drift requiring model retraining.
Generally: PSI < 0.1 is stable (no action needed), 0.1–0.25 is moderate drift (monitor closely), and PSI > 0.25 is severe drift (retrain or investigate immediately). We flag all columns above 0.1 in the report.
Yes. POST your two base64-encoded CSVs to /api/v1/detect-drift. Parse the response's drifted_features array — if non-empty and severity is "severe", exit with a non-zero status to block the deployment.
Free tier supports files up to 2MB each. Starter allows up to 20MB per file. Pro supports up to 100MB. For very large datasets, sample representative subsets — 10,000 rows is typically sufficient for reliable PSI computation.