Online recruitment fraud has become one of the fastest-growing categories of cybercrime, with reported losses rising rapidly and generative AI lowering the cost of producing convincing fraudulent listings. Existing defences leave four gaps: contextual signals from outside the listing are rarely incorporated, probability calibration is largely ignored, explainability is acknowledged but not addressed, and end-to-end deployment as a usable tool remains theoretical. This dissertation presents a calibrated hybrid ensemble for explainable online recruitment fraud detection that addresses all four gaps. The architecture combines two TF-IDF text classifiers (LinearSVC and SGD), a Gradient Boosting metadata classifier, and a fine-tuned DistilBERT transformer, applying Platt scaling to linear classifiers, isotonic regression to the transformer, and isotonic regression at each blending stage. A Bayesian signal floor activates when deterministic fraud markers fire on postings that the ensemble underweights. A three-layer company verification module (domain heuristics, DNS/HTTPS/WHOIS network checks, and Companies House registry lookup) returns an independent verdict on the claimed employer. A bidirectional explanation engine combines TF-IDF coefficients, Gradient Boosting feature importances, and LIME on DistilBERT to surface evidence of fraud and legitimacy for the user. On the EMSCAD test set, the deployed system achieves strong classification performance and excellent probability calibration. A user study with active job seekers reported a substantial post-test increase in self-reported confidence and broad agreement that the explanation surface clearly justified each verdict. The system is deployed as a Flask web application, with the dissertation making explicit the limits that dataset currency and statistical power impose on EMSCAD-based evaluation.