An Interpretable Multi-Signal Scam Detection System Using Machine Learning & LLMs

Scam and phishing detection systems often rely either on rigid heuristic rules or opaque large language models. Heuristics lack generalization, while LLMs are costly and difficult to audit. This work presents a hybrid, interpretable scam detection pipeline that combines a feature-based supervised machine learning model, semantic analysis via an LLM safety model, and rule-based heuristics within a unified ensemble decision framework. The system achieves 94.64% accuracy, 94.62% F1, and 0.9812 ROC-AUC on a balanced benchmark of 1,400 messages, with a Brier Score of 0.049 confirming well-calibrated probability outputs