Product reviews in Bangla text pose major analytical challenges due to the insufficient availability of language resources and benchmarked datasets. This research presents an upgraded multi-label category-based sentiment analysis framework for Bangla based on a hybrid deep learning approach. Our methodology combines multiple deep learning architectures, including LSTM, BiLSTM, and GRU, along with GloVe and FastText word embeddings and the transformer model BanglaBERT. Our proposed weighted ensemble strategy effectively balances model predictions, resulting in improved classification accuracy. The framework achieves a weighted F1 score of 0.88, outperforming standard machine learning models such as Logistic Regression (LR), Support Vector Machine (SVM), Naive Bayes (NB), AdaBoost, Gradient Boosting (GB), and Random Forest (RF), while also exceeding individual deep learning baselines including CNN, LSTM, BiLSTM, GRU, and CNN+BiLSTM. This research establishes a strong solution for sentiment analysis in Bangla through scalable e-commerce customer feedback analysis that improves user satisfaction and enables better decision-making.