To address the scarcity of extreme event data in air transport modeling, researchers introduced TailBooster, a dual-layer generative framework designed for extreme value augmentation in mixed-type tabular records. The methodology employs a statistical layer that utilizes the interquartile range to identify extreme values, which are then used to train a dedicated Tabular Variational Autoencoder (VAE) to generate tail-concentrated synthetic data. To prevent the generation of operationally infeasible instances, a second deep learning layer performs autoencoder-based cleaning, filtering out records that violate the operational envelope established by historical data. Evaluated using US flight records, the framework demonstrated significant improvements in operational validity and utility compared to conventional synthetic data methods. Specifically, when applied to six regression algorithms, training with TailBooster-augmented data reduced the Mean Absolute Error (MAE) by 47–49% for extreme air time prediction and by 29–57% for extreme arrival delay prediction. The study concludes that TailBooster provides a model-agnostic, data-driven approach that achieves performance gains comparable to enriching real datasets with synthetic extremes, making it highly effective for domains where domain-specific rules are unavailable but extreme-event prediction is critical.