You could have at least used embedding layers for the categorical features. Stacking dense layers with AutoML doesn't do much. So the comparison ends up being "good baseline vs something nobody ever does".