Architecture of M-TOPnet, the deep learning pipeline used for the galaxy spectral analysis in this work. Our model accepts two inputs: a normalised, continuum-subtracted spectrum and a normalised continuum vector. The spectrum is processed through convolutional residual blocks to extract spectral features, while the continuum vector is encoded through a dense layer. Outputs from both branches, along with intermediate outputs from the line location task, are combined into a shared information layer. This shared layer feeds into task-specific branches consisting of MLPs to predict line locations, redshift, and normalised values of M_star, and SFR. See a discussion on the different components in Sect. 3.2.