This paper explores a novel approach to sentiment analysis in the Bengali language, specifically when written using the English alphabet (transliterated Bengali). Traditional sentiment analysis methods, such as Support Vector Machines (SVM), K-Nearest Neighbors (KNN), and various Natural Language Processing (NLP) algorithms, have been widely used in previous studies. However, this research aims to leverage transformer-based Large Language Models (LLMs) that have undergone pretraining, offering an alternative methodology for real-time sentiment prediction. The primary objective of this research is to enhance and optimize existing LLMs to accurately interpret emotions expressed in transliterated Bengali text. The proposed real-time sentiment analysis engine utilizes modern LLMs to effectively gauge emotions in Bengali sentences written using English characters. This research also delves into transfer learning techniques, model architecture enhancements, and speed optimization strategies to further refine the LLM’s sentiment prediction capabilities. By extending the use of LLMs to language-specific challenges, particularly for languages with non-English scripts, this study contributes to the advancement of NLP. The pre-trained LLM model used throughout this research is Llama2-7B. This model outperformed the current methods with an accuracy of 83.47% and a training duration of 40.02 min.