Build a defensible ML pipeline end-to-end

You receive a tabular dataset with features (numerical, categorical, and text) and a binary target. Build an end-to-end modeling pipeline that you can defend under interview pressure: (1) define the business objective and the exact optimization metric (e.g., PR-AUC if positives are rare) and a decision threshold tied to cost/lift; (2) split strategy (time-based if temporal, with nested CV); (3) preprocessing: imputation plans, leakage checks, rare-category handling, and high-cardinality encoding (e.g., target encoding with out-of-fold schemes); (4) train at least two model families (e.g., regularized logistic regression and gradient boosting), perform hyperparameter search, and compare with calibrated probabilities; (5) evaluate stability, fairness across regions/job_category, and calibration; (6) produce feature importance with a model-agnostic method and a plan for monitoring in production (data drift, PSI, threshold re-tuning).