However, tooling for deep learning in general is not ready for industry grade technology. Many bugs could be prevented by dependent types, but compilers are not there. Also, debugging models feels like alchemy and random changes until it works. In addition, in production systems, rigorous testing is not a standard. The closest thing I have heard of is Tesla's data engine and AI system, they do have unit tests and a shadow mode. Of course big companies will have similar technologies for critical systems, but it's not as standardized as testing in software engineering.
Thanks for the contributions everyone, and glad to have bet on PyTorch a few years back.