Efficient electrophysiology simulation is critical for cardiac digital twins, yet clinical adoption remains limited by computational costs. Existing neural surrogate frameworks offer fast inference but are restricted to scalar activation-time prediction on single-ventricle geometries, discarding the action potential morphology that shapes ECG waveforms. To the best of our knowledge, no deep learning method has directly mapped patient-specific biventricular anatomy to full transmembrane voltage dynamics and onward to 12-lead ECG. In this study, we propose Geo-DONet, a geometry-parameterized deep operator network extended to 12-lead ECG prediction via a temporal-convolutional block. Validated on 125 biventricular geometries against the monodomain formulation, our model achieved a mean absolute error for transmembrane potential of 3.64~mV and a Pearson correlation of 0.83 for 12-lead ECG prediction, with an inference time of 290~ms vs 54~minutes on 128 cores for traditional solvers. This framework provides a scalable foundation for real-time, patient-specific EP simulation.