Deploy & close the loop
Use field events and evaluation results to decide what to inspect, label, and measure in the next data release.
Deployment adds new evidence to the data workflow. Field recordings and evaluation results show where the model succeeds, where it fails, and which conditions were missing from the training set.
The next step is a data decision, not an automatic retraining promise. Choose which events deserve investigation and preserve enough context to understand them.
Capture the useful context
Your fleet and evaluation systems determine what is recorded and when. Keep the device or run identifier, model version, sensor timestamps, calibration context, and reason an event was selected. Where your workspace has fleet integrations or recording rules configured, use them to organize the incoming data.
A confidence drop, a disengagement, or an unsuccessful grasp may identify a candidate event. Review it before assuming that labeling more of the same data is the right fix.
Route an event into a workflow
A practical sequence is:
event selected → recording inspected → labeling scope defined → labels created → quality reviewed → data exported → model evaluated
Decide which steps need human approval and which can be automated through available APIs and integrations. The process depends on your project configuration, data volume, quality requirements, and training infrastructure. It does not have a guaranteed completion time.
Check that the change helped
Evaluate the new model on the original failure scenario and on a broader regression set. A targeted improvement can hide a loss elsewhere. Keep the link from the evaluation result back to the data release so the outcome can inform the next decision.
Deploy through your existing release process once the relevant gates pass. Capturing an event, creating an annotation task, retraining, and deploying are distinct operations; connecting them requires deliberate orchestration.
Carry the evidence into the next turn
The loop is useful when a team can answer: what happened, what data changed, how was it reviewed, and what did the next evaluation show? Avala connects the data work that supports those answers.
The final lesson points to the SDK, CLI, and agent integrations you can use to connect that work to your systems.
Next: Make it yours →