Reduce overfitting under constraints

The model is overfitting: training RMSE is 4000 while validation RMSE is 9500. No additional data can be collected, and online inference has to remain below 20 ms at p95. Select and rank three interventions for reducing overfitting, describe how each one works, and outline an experiment plan; possible options include L1/L2/elastic-net regularization and their expected effect on coefficients, early stopping with patience, reducing architecture size or tree depth, feature selection or smoothed target encoding, tabular-data-appropriate augmentation, stratified K-fold cross-validation, bagging versus boosting, and leakage checks. Give concrete hyperparameter grids, metrics to monitor, stopping rules, and the method you would use to demonstrate statistically significant improvement.

Overview: The question assesses a candidate's ability to apply machine learning techniques to reduce overfitting in tabular regression while respecting production latency limits, covering regularization, model complexity control, feature engineering, validation strategies, and experiment design.