Now in beta

Your model is drifting.
You won't know until it's too late.

ModelSentry monitors your production ML models for data drift — locally, with no data ever leaving your machine.

Free for beta users. No credit card. No enterprise sales process.
The problem

Sound familiar?

Six months after you deployed the churn model, something changed. New marketing channels brought in a different demographic — younger users, different income brackets, different behavior patterns. The model kept predicting. Confidently. Silently. On data it had never seen.

You found out when a product manager pinged you asking why churn had jumped 12% in Q3.

You pulled the feature distributions. The age column had shifted by nearly 15 years. Income looked nothing like training. The model had been extrapolating for months — and you had no early warning system in place.

"There was no alarm. There should have been."

How it works

Three steps. No infrastructure.

Step 1

Install (30 seconds)

pip install modelsentry

Runs anywhere Python runs. No Docker. No cloud account. No infra to manage.

Step 2

Wrap your predict function

import modelsentry as ms ms.init(model_id="churn-v3", profile_window=500) @ms.monitor() def predict(features_df): return model.predict(features_df)

One decorator. No changes to your model logic. No SDK calls inside predict().

Step 3

Open the dashboard

modelsentry serve \ --model churn-v3

Opens localhost:8080. PSI and KS drift scores update as predictions roll in. You get an email the moment drift crosses your threshold.

The dashboard

Tells you what changed, not just that something did.

Most monitoring tools tell you "accuracy dropped." ModelSentry tells you which features drifted, by how much, and whether it's worth waking up for.

  • Per-feature PSI and KS statistics, color-coded by severity
  • Baseline vs. current distribution comparison for every feature
  • Alert history with timestamps — know when drift first appeared
  • Prediction volume counter — catch traffic shifts too
  • Auto-refreshes every 60 seconds while the tab is open
churn-v3 ■ Critical drift
Feature PSI KS p-val Severity
age 0.31 0.001 ■ critical
income 0.18 0.023 ▲ warning
country 0.04 0.412 ● stable
Last updated: 2 minutes ago Predictions: 12,847

Your data never leaves your machine.

ModelSentry computes statistical profiles locally — histograms, means, PSI scores, KS statistics. Raw feature values and raw predictions are never written to disk, never transmitted over a network, never seen by anyone.

The dashboard runs at localhost:8080. Email alerts are sent from your own Gmail account via your own app password. We have no servers that touch your data. There is nothing to breach.

Raw feature values — never persisted, never transmitted
Raw predictions — never written to disk
Dashboard — localhost only, never exposed to the network
SMTP — your credentials, your inbox, your control
Early access

Get early access.

We're opening the beta to a small group of data scientists. No credit card. No enterprise sales process. No SaaS subscription to cancel.

Install, monitor, and tell us what breaks.

Free for beta users, forever.

No credit card. Takes 30 seconds.