I agree with this. In my experience, LightGBM is often good enough. The additional improvement provided by neural net models is not worth the substantial increase in development time.