Every day, e-commerce platforms gather vast numbers of textual customer reviews that reflect detailed evaluations of products and services. Sentiment analysis techniques from previous traditional frameworks often struggle to detect the complex emotional aspects in these instances, particularly within multi-component evaluation contexts. To address these challenges, we propose Insight Orbit, a refined Natural Language Processing (NLP) framework that performs aspect-based sentiment analysis on product reviews while providing interpretable justifications via a generative model. Specifically, we utilized Bidirectional Encoder Representations from Transformers (BERT and DistilBERT) for aspect extraction and sentiment analysis and integrated it with GPT-2 to generate sentiment scores along with coherent justifications for each aspect. This enhanced version of the Insight Orbit system addresses several key limitations previously identified in the literature and our preliminary work, including (1) more robust data preprocessing and bias detection, (2) a thorough error and statistical significance analysis, (3) deeper user-centric insights facilitated by more transparent justifications, and (4) a scalable Azure Functions deployment for real-time, production-ready sentiment classification. Experimental results on a large Amazon product review dataset demonstrate that our transformer-based approach outperforms classical methods—yielding an accuracy of 94.5% (DistilBERT) and 95.2% (BERT) (with corresponding F1 scores of 0.945 and 0.951, respectively) versus 88.9% accuracy (F1 = 0.889) for a Logistic Regression baseline. Latency measurements indicate average end-to-end processing times of 150–200 ms under normal loads and up to 350 ms during peak concurrent requests. These improvements are further confirmed by the use of statistical significance tests (p < 0.01).