TheTutor

Ask most AI tools to write a syllabus and they will — fluently, confidently, and from nowhere in particular. It’s the model’s general sense of a subject, not any specific book, which is fine until a learner asks where a claim came from and there’s no answer. TheTutor’s starting constraint was that every syllabus it writes has to point back to a real textbook, not the model’s memory of one.

That meant treating the textbook as something the model has to go and read, not something folded into a prompt. MCP is the wiring for that: the generation step calls out to the textbook corpus as a tool mid-generation, gets back an actual passage, and writes the course section from what it was just handed — so a citation traces to a retrieved page rather than to a guess the model is confident about. Next.js and Express carry the generation pipeline and the admin/course tooling around it; MongoDB Atlas holds the courses, the retrieved sources and the links between them.

The constraint that shaped the build more than the tech stack did was the calendar: five engineers, three weeks, ~22,400 lines of TypeScript. Grounding is the one piece of that pipeline that can’t be handwaved under time pressure — a course generator that occasionally cites the wrong page is worse than one that admits it found nothing — so that path got the most scrutiny of anything shipped, and the rest of the build moved fast around it rather than through it.