The LEGEND experiment searches for neutrinoless double-beta decay using high-purity germanium detectors. Background rejection relies heavily on pulse shape discrimination, traditionally using the A/E method. A previous work (Holl et al., Eur. Phys. J. C (2019) 79:450) demonstrated that a simple two-stage neural network can match A/E performance on BEGe detectors while capturing background events that A/E misses. In this talk, I present an initial attempt to apply this architecture to LEGEND calibration data. I will also discuss practical lessons learned, particularly the need for efficient waveform data access in the current LEGEND software stack.