Catch drift before your users do
Continuous statistical monitoring on every inference request. Multiple drift metrics evaluated against a stable training baseline. Alerts before the SLA breaks — not after.
Three types of drift. One monitor.
Different failure modes require different statistical tests. Inferpathio runs all of them in parallel and surfaces the right signal.
Feature distribution drift
Compare how input feature distributions shift between training baseline and production windows. PSI < 0.1 is stable; > 0.2 triggers an alert.
Data shift detection
Non-parametric test for whether two samples come from the same distribution. Catches subtle continuous-feature drift before PSI does.
Prediction output drift
Earth mover's distance between prediction score distributions. Catches model behavior changes even when input features look stable.
Alerts that go where your team works
Configure thresholds per model, per feature group. Route alerts to Slack, PagerDuty, webhook, or directly trigger a retraining pipeline. No manual check-in required.
model: revenue-forecast-v3
baseline: run-7291
metrics:
- psi: threshold=0.10
- ks_test: p_value=0.05
- wasserstein: threshold=0.15
on_trigger:
- slack: "#ml-ops-alerts"
- action: retrain
pipeline: revenue-retrain-v2Stop monitoring manually
Set up your first drift monitor in 10 minutes on the free plan.