Back office ops · Production
Building a continuous ML training-serving pipeline with Vertex AI and Superwise
The problem
ML models in production degrade silently due to constantly evolving data profiles, and teams lack infrastructure to automatically detect this degradation and trigger retraining before the model fails.
Workflow diagram · grounded in source
1
Distribution shift triggers retraining
trigger
“training-serving pipeline in place that will be initiated in the event of a distribution shift in the production data”
2
Load and validate dataset
integration
“Extract the Diamonds dataset, and use only the >10,000 priced diamonds for training. Validate the dataset – some simple feature engineering.”
3
Train RandomForestRegressor
ai_action
“Train a RandomForestRegressor on the training data.”
4
Evaluate and validate model
validation
“Evaluate our model on the test data. Validate that the model is production-ready.”
5
Register model to Superwise
integration
“Generate a model and a version in our monitoring system.”
6
Deploy to Vertex endpoint
output
“Deploy our model to an endpoint.”
7
Log predictions for drift monitoring
feedback_loop
“_send_monitor_data – for each prediction request send the prediction data to Superwise”
Reported outcome
The build demonstrates a training-serving pipeline that automatically retrains when production data distribution shifts are detected, with each new model version registered in Superwise for ongoing monitoring.
Reported stack
VertexKubeflowSuperwiseGoogle storageGoogle cloud functionFlaskRandomForestRegressorscikit-learn
Frequently asked questions
What did this team achieve with this AI workflow?
The build demonstrates a training-serving pipeline that automatically retrains when production data distribution shifts are detected, with each new model version registered in Superwise for ongoing monitoring.
What tools did this team use?
Vertex, Kubeflow, Superwise, Google storage, Google cloud function, Flask, RandomForestRegressor, scikit-learn.
How is this back office ops AI workflow structured?
Distribution shift triggers retraining → Load and validate dataset → Train RandomForestRegressor → Evaluate and validate model → Register model to Superwise → Deploy to Vertex endpoint → Log predictions for drift monitoring.