This paper presents SpamShield, a production-grade, modular spam detection system that unifies three classifier paradigms—Multinomial Naive Bayes (MNB) within a fully encapsulated sklearn Pipeline, Support Vector Machine (SVM) optimised via GridSearchCV across a nine-combination hyperparameter grid, and a five-layer Deep Neural Network (DNN) with Batch Normalisation, progressive Dropout regularisation, and class-weighted training—with Local Interpretable Model-agnostic Explanations (LIME) post-hoc explainability. Evaluated on the UCI SMS Spam Collection (5,571 messages, 13.4% spam), the system achieves DNN accuracy of 99.2%, F1=98.9%, and AUC=0.999 with only 5 false positives on 1,115 test messages. The GridSearchCV-optimised SVM achieves 98.5% accuracy and F1=98.3% with 22 false positives, providing a practical high-performance alternative. A novel multi-class sklearn Pipeline replaces hardcoded regex category classification, achieving 94.0% accuracy across five spam sub-categories (Financial, Promotional, Scam, Adult, Phishing). The system features sklearn Pipeline encapsulation to eliminate preprocessing skew, rotating file logging, a three-tier Flask Blueprint architecture, real-time Gmail API integration with token-bucket rate limiting and automated spam-folder routing, LIME explanations with vocabulary learned from NB log-probability differences, and a fully dynamic frontend with all metrics fetched from REST APIs. This work demonstrates that architectural discipline—modular code, full Pipeline encapsulation, systematic hyperparameter search, and explainability integration—transforms a prototype classifier into a portfolio-grade production system.

Index Terms— email classification, hyperparameter optimization, interpretable artificial intelligence, automated message filtering, production-ready machine learning