I think deep and learning is the key part of this topic. A practical approach is to define the outcome you want, test it with a small example, and then check quality, speed and reliability separately. The original point — "When working with Deep Learning, the biggest gains usually come from...