I'd recommend moving to development that interfaces with ML instead of moving to being an ML practitioner. You have much faster feedback cycles, your work is predictable (engineering, not science), and you don't feel the pressure of never reading enough papers.

To do that you only need to understand the fundamentals of tensors, some basic knowledge on what the big no-nos are within ML development so you can course correct your peers if they break them, and either focus on the operations side of things or deployment. In both cases, having a knack for optimizing bottlenecks will be very helpful since they will be present during both training and inference.