Pradra StudioDocs
Sign in

Monitoring

Shipping a model is the start, not the finish. Once a capability is deployed, Pradra Studio watches it on several fronts — so the first person to know about a problem is you, not your users.

What is watched

  • Operational health — how fast the endpoint responds (with latency percentiles, not just averages), how often it errors, how much traffic it handles. This tells you the service is healthy.
  • Data drift — whether the inputs arriving in production have moved away from the data the model was trained on. Covered in depth in Drift & retraining.
  • Model behavior — whether the model's own predictions are shifting compared to what it learned.
  • Cost — for generative capabilities, token usage and spend per call. See Budgets & costs.

Where to look in the Studio

  1. Monitoring in the sidebar is the fleet view: every live deployment in the workspace, its request rate, error rate, and latency at a glance — the page you keep open on the ops screen.
  2. Each deployment's detail page is the deep dive: request and error charts over time, latency percentiles, drift status, and the prediction history feeding it all.
  3. The Dashboard gives the cross-cutting summary — resource counts, latest deployments with health dots, recent training runs.

Alerts fire when a metric crosses a threshold — sustained error rates, latency spikes, drift scores, silent drift evaluation — so a quiet dashboard means something.

An honest signal, by design

Pradra Studio is careful never to fake confidence:

  • When there is too little traffic to judge drift, the platform reports "not enough evidence" — it does not report a score of zero and pretend all is well.
  • A quiet endpoint is flagged as quiet, not as healthy.
  • A monitoring gap (an evaluation that stopped running) raises its own alert, rather than leaving a stale green light.

This is deliberate: an alert you can trust is worth more than a dashboard that is always green.

Quality over time

Operational metrics tell you the service works; they don't tell you the predictions are still right. Judging accuracy in production needs real outcomes (did the customer actually churn?), which arrive later, if at all. Until they do, data drift is your early-warning system — if the world the model sees has changed, its accuracy is in question even before you can measure it.

Next