Decades of cognitive science tell us what helps students learn — the testing effect, spaced repetition, elaboration, anchoring, and targeted feedback. The problem is that traditional ed tech tools struggle to support those practices at scale. Large language models change that equation. This presentation explores a working AI tutor built around Harald Prins's Barrel Model of Culture — a system that brings genuine individualized learning to every student by embedding these evidence-based principles into an adaptive, conversational experience. We'll examine the design decisions that make it work and what they reveal about building AI tools grounded in learning science, not novelty. Attendees leave with concrete patterns they can apply immediately.