Eight years of algorithms work in a defense R&D environment, before I ever touched a product roadmap, shaped how I think about shipping AI more than anything I've done since. The standard there wasn't "does it work in a notebook" — it was "does it work when the input distribution shifts under real conditions, and can you prove it."
That rigor is easy to lose in a fast-moving product company, where the pressure is to ship the demo and iterate later. I carried the habit forward anyway: test against adversarial and edge-case inputs before you trust a result, and treat "it passed on our validation set" as the start of the evaluation, not the end.
It also taught me to separate the algorithm from the system around it. A good model wrapped in a fragile pipeline fails in production for reasons that have nothing to do with the model. Most of what I now call "AI governance" is really that same lesson, generalized: the system around the model is where trust is actually won or lost.