Do machine learning engineers ever implement neural networks from scratch for any given problems?

Mostly no. The architecture of neural networks (be it shallow or deep) might be designed from scratch depending on the problem at hand but the actual implementation will be done using available libraries.

P.S: From my experience, the only time people would have implemented neural networks from scratch would be when they were learning it for the first time.