Zynth

Inspiration## Inspiration

I can spend three hours on a chapter and still not know whether I can actually do any of it. Most study apps either explain things at you or quiz you. Neither one tells you what you actually know.

What it does

Zynth builds your syllabus into a 3D knowledge graph. Each concept is a node, and the colour means one thing: proof.

Red is untouched. Reading the chapter nine times won't change it. Amber means you've engaged with it but nobody has checked. Green means you passed a quiz, and that's the only way in.

I put that rule in the database rather than the app, so a SQLite trigger throws out any illegal change. You can open a SQL console, try to set a red node to green, and it just refuses. Green isn't permanent either. Fail a retest and it drops back to amber.

The part I like most is the Autopsy Board. You paste in your wrong answers and it works out the one misconception sitting under all of them, then draws new edges between the concepts that keep breaking together. I fed it four mistakes from Computer Science, Physics and AI. It worked out I kept naming a system's input when the question asked for its output, and connected all three subjects on the map.

How I built it

React and Three.js on the front, Node with SQLite and Socket.io behind it. Gemini runs the agents. Groq grades the written answers, deliberately a different model from the one that set the question.

The rule I stuck to: the models write and diagnose, the code decides. Nothing Gemini says can turn a node green.

Challenges

Getting the Co-Pilot to stay quiet. If it fires on every wrong answer it's just noise and you learn to ignore it. I threw 17 awkward scenarios at the detection logic before wiring any of it in, and ended up rewriting the rules twice because the scenarios kept proving them wrong. Shutting it up was harder than getting it to notice things.

Accomplishments

You can't cheat the status rule. I tried eight illegal transitions and a replay attack, and all of them bounced. A red node that answers everything correctly still stays red, because it skipped the step before.

What I learned

Most of the work went into deciding when the AI should say nothing. Getting a model to produce an insight is easy. Working out whether that insight is worth interrupting someone mid-quiz is the part that took time.

What's next

Getting it in front of people at school. I want to see whether opening on a screen full of red makes you want to fix it or makes you close the tab. I think the first. Could be wrong.