Drift Monitor

Live model monitoring

A simulated process emits observations around hidden class centres. A PyTorch classifier predicts each batch; when accuracy stays below the threshold, it retrains on the recent window and redeploys.

Live accuracy
—
Model state
Starting
Training the initial model
Deployed model
—
Validation —
Interval t
—
Session —
Data drift
None
Centres stable

Live accuracy per interval

Share of each new batch the deployed model classifies correctly. Shaded spans mark data drift; vertical rules mark redeployments.

Waiting for the first interval…

Data in 2-D projection

Loading…

  • P k current centre

Controls

Trigger drift, retrain, and tune the monitoring policy.

The dashboard is public; controls need an operator session. Sign in

Training runs · validation loss per epoch

Validation is the most recent intervals of the training window. Training stops when validation loss stops improving; the dot marks the epoch whose weights were deployed.

No training runs yet…

Training parameters

AdamW with early stopping. Changes apply from the next training.

Loading…

Event log

Newest first. Every training run is also logged to MLflow.

No events yet.