Bangla serves as one of the most commonly spoken languages because millions of people use it to share their political and social views through online platforms. Natural language processing for Bangla remains underdeveloped because its complex morphological rules and spelling differences and common codemixing practices and shortage of quality annotated data create challenges for political sentiment analysis. The research examines more than 9000 Bangla political comments which have been manually annotated to show three different sentiment categories of positive and negative and neutral. The unified preprocessing pipeline uses text normalization and duplicate conflict resolution and lexicon-based label correction and class imbalance handling to fix data quality problems which allow for unbiased assessment. The research tests common machine learning techniques and deep learning models and transformer-based systems through one experimental setup which uses hybrid TF-IDF features and fastText word embeddings and subword-based contextual representations. The experimental results show that transformerbased methods provide better performance than traditional and recurrent approaches while the combination of BanglaBERT XLM-RoBERTa and multilingual BERT achieves 88.19 % accuracy with 88.16% F1-score. The study proves that large language models successfully capture contextual and semantic details of Bangla political discourse which creates a research benchmark for low-resource language sentiment analysis studies.