AIONIS Maieutics

Inspiration

Every AI tutor I tried did the same thing: I asked, it answered. I felt smart for a second and forgot everything by the next day. The problem isn't that students lack access to answers, it's that being handed one skips the only part that actually builds understanding: reconstructing the idea yourself. I wanted a tutor that refuses to do that work for you.

What it does

AIONIS Maieutics is a Socratic tutor that never gives you the answer. You pick a STEM concept, and it guides you with questions only, until you rebuild the concept from first principles in your own words. As you reconstruct each piece, a concept graph lights up node by node next to the conversation. If you contradict a foundation you already built, that piece is gently flagged so you can revisit it, never punished. The engine is topic-blind: the same code guides you through gas pressure, gradient descent, predator-prey cycles, and Braess's paradox.

How we built it

Next.js, with the engine running entirely server-side behind route handlers, so the answer key never reaches the browser. The core is a topic-blind state machine over a concept dependency graph, with a single structured LLM call per turn that classifies whether you rebuilt the mechanism, guessed, or broke a prerequisite. Claude Sonnet runs the evaluation, with dual-provider failover so a live demo never dies mid-session. The graph view uses React Flow, deployed on Railway.

Challenges we ran into

The hardest part wasn't the code, it was the feel. My first version was technically correct but pedantic: it took six turns to accept a right answer, and it kept dragging me back to steps I'd already mastered. I literally typed "I ALREADY TOLD YOU" three times during a test. Fixing that without breaking the soul, the refusal to ever hand over the answer, was the real work: lowering the bar to accept a clear mechanism even when poorly worded, while still rejecting bare term-naming, plus an iron rule to never retreat to an already-rebuilt node.

Accomplishments that we're proud of

That the engine is genuinely topic-blind, proven not claimed: the exact same code reconstructs physics, machine learning, and ecology with zero changes. And that the same structural primitives, like FEEDBACK and EMERGENCE, reappear across unrelated domains, the seed of seeing the world's structure as a shared language.

What we learned

That a good Socratic tutor is patient, not stubborn. Refusing to give the answer isn't the same as leaving someone stuck against a wall. The whole craft was teaching it to lower the difficulty step by step until it finds a foothold you can climb, without ever revealing the destination.

What's next for AIONIS Maieutics

Right now AIONIS teaches each concept as a specific instance. The real vision is the abstract pattern beneath it: Braess's paradox isn't only about traffic, it's a mechanism that reappears in power grids and markets. The next step is a tutor that shows you the pattern itself, crossing domains, turning the structure of the world into a language you can read.