Sentiment analysis is the process of identifying and categorizing opinions expressed in a piece of text. It has been extensively studied for languages like English and Chinese but still needs to be explored for languages such as Urdu and Hindi. This paper presents an in-depth analysis of Urdu text using state-of-the-art supervised learning techniques and a transformer-based technique. We manually annotated and preprocessed the dataset from various Urdu blog websites to categorize the sentiments into positive, neutral, and negative classes. We utilize five machine learning classifiers: Support Vector Machine (SVM), K-nearest neighbor (KNN), Naive Bayes, Multinomial Logistic Regression (MLR), and the transformer-based multilingual BERT (mBERT) model. This model was fine-tuned to capture deep contextual embeddings specific to Urdu text. The mBERT model was pre-trained on 104 languages and optimized for Urdu-specific sentiment classification by fine-tuning it on the dataset. Our results demonstrated that the mBERT model significantly outperformed traditional classifiers, achieving an accuracy of 96.5% on the test set. The study highlights the effectiveness of transfer learning via mBERT for low-resource languages such as Urdu, making it a highly promising approach for sentiment analysis.