Built a hybrid sentiment model using BERT-based augmentation, transformer embeddings, ensemble learners, and a CNN-BiGRU-Attention network. Reached 91.15% accuracy on Amazon reviews via stacking and Optuna hyperparameter tuning. Used contextual augmentation, mpnet embeddings, and SMOTE to improve class balance and data quality