The context-dependence of sarcasm recognition in textual data makes it a challenging problem. This article describes a fusion technique that combines the XGBoost gradient boosting algorithm with BERT embeddings for sarcasm detection. Using this method, textual input is transformed into rich contextual embeddings that are then used to train an XGBoost classifier. The resulting hybrid model is evaluated and trained across multiple datasets, demonstrating its ability to distinguish between text that is sarcastic and non-sarcastic. The results of the experiment indicate that the accuracy matrix performs better than individual models. A comprehensive word cloud analysis that identifies important phrases and language patterns associated with sarcasm is also included in this paper. This study combines the benefits of BERT and XGBoost to advance sarcasm detection techniques.