START

Initialize System

Load Models (MLflow)

Prod (champion)+ 2 Challengers

(10-second batches)

100 samples

Generate New Data + Scale

Calculate Metrics

Accuracy, Drift, Latency

Print Summary

>= 10?

(last 10 batches)

Bayesian Optimization

Model in MLflow

5-Batch Evaluation

(count: X/5)

5 batches?

>= 60%?

< 3?

(new params)

Max Retries Reached

Parallel Inference

Simultaneous

Prod Model

(Champion)

All Challengers