DRIFTWATCH

LLM production console — a model drift & MLOps teaching simulation

WEEK 0

Live eval accuracy weekly automated eval against fresh, human-labeled traffic

Eval accuracy Degraded (<88%) SLA breach (<75%) Deployment

Input distribution: training snapshot vs. live traffic the gap between the curves is the drift

What the model was trained on What users ask about now Divergence (unfamiliar territory)

Service health

Eval accuracy
Drift score (PSI)
threshold 0.25
Model age
weeks since deploy
Retrain spend
$0
0 pipeline runs

MLOps cycle

Monitoring production…

Ops controls

0.030
0.25

Live traffic sample ✓ answered correctly · ✗ stale answer

Ops event log

Why accuracy decays

The model's weights are frozen at deployment, but the world keeps moving: new products, new APIs, new slang, new regulations. The blue curve (training data) stays put while the teal curve (live traffic) slides away. The bigger the gap, the more often the model answers from a world that no longer exists.

Why we monitor drift, not just accuracy

Accuracy needs labeled data, which is slow and expensive to collect. The drift score (PSI) compares input distributions only — no labels needed — so it acts as an early-warning smoke detector. Watch how the PSI alert usually fires before accuracy visibly craters.

Why retraining never fully catches up

Notice that a fresh deployment starts slightly behind: the pipeline trains on data collected when it started, and the world moved during training. That's the MLOps treadmill — monitor → collect → train → evaluate → deploy → monitor — and why teams automate it instead of retraining by hand.