The research introduces a new framework to analyze sentiment across multiple languages which aims to boost customer satisfaction on international e-commerce websites namely Amazon, Walmart, Apple, eBay, and Etsy. Multilingual BERT (mBERT) processed a dataset of 13,000 customer reviews spanning English and Hindi and their code-mixed variants along with German and Spanish texts. It then tokenized and embedded the texts for subsequent analysis. We created a combined architecture that uses mBERT embeddings together with an Attention-Augmented GRU layer to handle linguistic uniqueness and situational ambiguities. The proposed method links transformer-based global context understanding with GRUs built for modeling sequence relationships while adding attention mechanics to highlight significant textual elements. The proposed model reached 93.45% test accuracy and 0.0974 as test loss which established superior performance compared to standard architectures such as LSTM, BiLSTM and BiLSTM-GRU. The framework delivers an adaptable blueprint for sentiment analysis improvement across multiple linguistic contexts by resolving tokenization issues while using advanced deep learning techniques which enhance customer satisfaction accuracy.