Email volume continues to surge across business, education, and personal use, with spam composing nearly half of global traffic and driving financial and security risks. Classical filters (blacklists, rules, Naive Bayes, k-Nearest Neighbors) and even deep models like BiLSTM struggle to capture full semantics or resist modern evasion. We propose a BERT-driven spam-detection framework that exploits bidirectional self-attention to model rich context and evaluate it against strong baselines. On a balanced dataset of 5,000 emails with a 226-sample hold-out test set, fine-tuned BERT attains 98.67% accuracy and 98.65% F1, outperforming BiLSTM (96.43%), Naive Bayes (91.59%), and k-NN (90.27%). For practical deployment, we integrate BERT into a Chrome extension with a FastAPI backend, enabling real-time Gmail filtering with 0.5 seconds latency per email. Results indicate BERT delivers superior accuracy and robustness, bridging research and production. Future directions include multilingual detection, adversarially robust training, and multimodal filtering (URLs, images).