Sentiment analysis of large-scale e-commerce review data presents significant challenges, particularly due to imbalanced class distributions and the need for deep semantic understanding. In this work, we propose a robust hybrid pipeline that integrates BGE (BAAI General Embedding) sentence embeddings with a Rainbow Deep Q-Network (DQN) reinforcement learning framework for binary sentiment classification. The system frames sentiment prediction as a reward-driven decision problem, where the agent learns to classify reviews as Positive or Negative by receiving class-weighted rewards. The pipeline preprocesses Flipkart customer reviews, generates normalized BGE-small-en embeddings, and trains a Rainbow DQN with dueling architecture, Adam optimizer, and Huber loss. Experiments on a subset of 50,000 reviews demonstrate that the proposed method achieves an overall accuracy of 96.19%, with per-class F1-scores of 0.98 (Positive) and 0.88 (Negative), outperforming traditional ML and deep learning baselines. The study highlights the effectiveness of combining semantic embeddings with adaptive reinforcement learning for scalable, imbalance-aware sentiment classification.