Everyone's demoed AI writing a function. Far fewer have made AI a reliable part of shipping production backend code. This talk is the honest, experience-based account of where LLMs help, where they fail expensively, and the guardrails that turn them from a toy into a force multiplier.

A realistic model of AI as an *augmented assistant behind guardrails* rather than a fire-and-forget generator, with measurable criteria for where to deploy it in their own flow.