Sentiment analysis, an important part of natural language processing, looks for opinions and emotional polarity in written data like product reviews, social media posts, and customer feedback. Organizations can better understand user perspectives and make well-informed decisions with the aid of accurate sentiment classification. In order to enhance classification performance, we present an ensemble-based sentiment analysis framework in this work that combines transformer architectures with deep learning methods. The suggested method captures both contextual and hierarchical relationships in textual data using a Capsule Neural Network (CapsNet) and pre-trained transformer models like RoBERTa and ELECTRA. The models undergo preprocessing, tokenization, and evaluation on the SST-2 sentiment dataset.as well as feature extraction. A weighted ensemble voting strategy is used to combine the outputs of individual models based on optimized weight configurations in order to improve prediction accuracy and stability. According to experimental results, the accuracy of the RoBERTa model was 93.55%, whereas the accuracy of ELECTRA and CapsNet was 94.70%. The weighted ensemble model demonstrated consistent and dependable sentiment classification performance with an overall accuracy and F1- score of 94.70%. The suggested framework demonstrates how transformer-based architectures can be successfully combined with capsule networks and ensemble learning methods to create reliable sentiment analysis systems.