Follow the machine learning lifecycle — decision exercise
Exercise: review a manufacturing-model handover
Original fictional case. No paid services or cloud account required. Difficulty: beginner · Estimated duration: 15 minutes
A team collected inspection records, corrected inconsistent labels, trained version A and evaluated it on held-out cases. Version B exists, but the handover does not say which version the application serves. Production inputs changed last week and no follow-up evaluation is recorded.
| Evidence | State |
|---|---|
| Data collection and label repair | Recorded |
| Version A evaluation | Recorded |
| Serving version and configuration | Missing |
| Evaluation of changed inputs | Missing |
Your tasks
- Map the recorded and missing evidence to ingestion, preparation, training, deployment and management.
- Explain what Cloud Storage, Model Garden, Model Registry and an AutoML path can contribute. Do not claim any tool alone establishes readiness.
- Recommend the next handover checks without discarding the useful historical evidence.
Reference solution
Collection is ingestion; label repair is preparation; training produced version A; its held-out results are evaluation evidence. The intended serving version and configuration still need verification, and management needs to investigate changed inputs and reevaluate relevant behavior. Cloud Storage can hold the input objects in buckets; storing objects is not training. Model Garden helps discover/test candidates. Model Registry helps organize versions and their lifecycle. AutoML can build supported model types from supplied training data without custom model code; the case does not require retraining merely because AutoML exists. Obtain the serving identity and configuration, connect the evaluation to that version and workload, then assign the next operational checks. Neither renaming a version nor assuming B is better supplies missing evidence.
Rubric (6 house points)
Two points each for: lifecycle mapping; distinct tool roles; an evidence-based next step that separates historical evaluation from current readiness.
Sources
- https://docs.cloud.google.com/architecture/deploy-operate-generative-ai-applications
- https://docs.cloud.google.com/gemini-enterprise-agent-platform/machine-learning/model-registry/introduction
- https://developers.google.com/machine-learning/glossary
- https://docs.cloud.google.com/gemini-enterprise-agent-platform/machine-learning/training-overview
- https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/model-garden/explore-models
- https://docs.cloud.google.com/storage/docs/introduction