Designing AI-Powered Learning Systems
Start from the learner, not the model
The first version of every question was: does this help a student understand a competence faster? Features like 'AI magic' that failed that question got cut. Dashboards, quizzes and document access did most of the heavy lifting before any model call was made.
Where AI genuinely helped
Three places earned their keep: answering questions grounded in the student's own materials, generating revision variants from existing content, and explaining a wrong quiz answer in context. In each case the AI was amplifying content the platform already had, not replacing it.
Where ordinary engineering helped more
Progress tracking, role management for teachers and parents, content versioning — none of that needs a model. Keeping those layers deterministic made the AI layer easier to trust, because the data going in and the audit coming out were exact.
Assessment integrity
An AI tutor next to an assessment system is a design tension. The split that worked: the tutor can explain, hint and revise, but assessment sessions run against a locked context with no tutor access. Learners get help while they learn, not while they are being measured.
The loop that matters
Watch where students actually struggle → change the smallest piece of the system that fixes it → measure again. Sometimes the fix is a better retrieval filter, sometimes it is one more sentence in a prompt, sometimes it is a plain UI change. Treating all three as the same engineering loop is the real lesson.