Solo founder project — 2025
A Socratic dialogue engine that will not give you the answer — grounded in a knowledge graph so it knows exactly what it is withholding and what to ask next.
In development
Image slot
AI Tutor — Socratic dialogue over a knowledge graph
1600×1000 · UI screenshot or a render of the concept graph
Role
Solo founder — concept, knowledge graph design, engine, front end
Year
2025
Stack
- Next.js
- TypeScript
- Knowledge graph
- LLM orchestration
The problem
An LLM asked to tutor will answer the question. That is what it is for, and it is precisely the wrong behaviour — the student gets a correct answer and no model of how to reach it. Prompting a model to 'be Socratic' produces the mannerisms of questioning without the substance, because the model has no representation of what the student does and does not yet hold.
Approach
Modelled the subject as a knowledge graph of concepts and dependencies, so the engine can locate a student's answer on a map rather than judging it in isolation.
Made the graph the source of the next question: given where the answer sits and which prerequisite it skipped, the next prompt targets that specific gap.
Held the engine to withholding — it can confirm, narrow, and redirect, but the final step stays with the student.
What it came to
Building it made clear that the hard problem is not the dialogue, it is the representation. Once the graph knew what depended on what, good questions mostly fell out of it.