Accurate localization of ventricular tachycardia (VT) exit sites is essential for effective ablation therapy, yet current pace mapping (PM) techniques provide limited, qualitative feedback and remain highly operator-dependent. We present the first real-time, data-driven framework that directly transforms standard 12-lead ECG signals into quantitative localization of VT exit sites and directional guidance for catheter navigation. By framing the problem as a linear inverse regression, we estimate spatial displacements from QRS integral features using a machine learning pipeline. Four regression models—multivariate, Ridge, Lasso, and Elastic Net—were evaluated on clinical data from four VT ablation patients. Sparse regression models, especially Elastic Net without intercept, achieved the best performance, yielding a mean localization error of 11.5 ± 3.1 mm using only 14 pacing sites. Our system continuously refines its prediction and generates real-time directional vectors to steer the catheter toward the VT exit site. This proof-of-concept study demonstrates that standard ECG can be repurposed into an intelligent, real-time navigation tool, potentially transforming clinical electrophysiology by improving precision and reducing reliance on operator experience.