Problem and data
Assess the decision and available data first. Missing values, measurement errors, class balance and time relationships precede modelling. Training on information unavailable at prediction time can create misleading results.
Development and integration
Preparation, feature extraction and model selection share an evaluation workflow. Compare against a simple baseline. API or application integration needs defined versions and input/output contracts.
Evaluation and deployment
Choose splits and metrics suited to the data. Future information must not leak into historical time-series tests. Monitor changing data and model errors after deployment; acceptance criteria relate to a specific operational need.