I wanted to start a broader mistakes and lessons thread around Full-Stack AI Development.
A useful angle is evaluation, but it also connects with security. Different tools and workflows can produce very different results even when the goal is similar.
It would be interesting to compare...