We used a shallow neural network. The main challenge at Sculpt wasn't the modelling part, but rather the whole UX (active learning to speed up the training proc

We used a shallow neural network. The main challenge at Sculpt wasn't the modelling part, but rather the whole UX (active learning to speed up the training process, explainability, easily get predictions, etc). So it's true that a relatively simple model performs well, and actually having an efficient / shallow network also helps make the active learning pipeline fast for the user.