MindMap Learning

Main interface
AI agent that talks

Inspiration

Most educational platforms measure the wrong thing.

They track time spent, videos watched, and quiz scores. But none of them measure misconceptions — the gaps that exist even when a student thinks they understand something.

A student says "I know recursion." A standard system asks for a definition. Score: 100%. But ask them what happens in memory during a recursive call and the understanding collapses.

That's the problem MindMap Learning solves.

What It Does

MindMap Learning is a Socratic AI examiner that:

Takes any topic a student claims to know
Interviews them with deep, reasoning-based questions (not MCQs)
Builds a live knowledge graph showing confidence vs competence for every concept
Detects the exact misconceptions — down to the specific node
Generates a personalized remediation plan targeting only the weak areas

The key insight: high confidence + low competence is the most dangerous learning state. MindMap makes that visible.

How I Built It

The app uses a 4-agent AI pipeline, each with a distinct role:

Agent	Role
Knowledge Mapper	Builds the concept graph for any topic
Socratic Examiner	Asks adaptive, reasoning-revealing questions
Misconception Detector	Scores confidence vs competence after each response
Remediation Coach	Generates a targeted learning plan at the end

Tech stack:

Frontend: React + Tailwind CSS (built with Lovable)
AI: Groq API (Llama 3 70B) — free, fast, no credits required
Architecture: Fully client-side, no backend needed
State: React useState managing the full knowledge graph and session

Challenges

Getting the agents to stay in JSON format was harder than expected. LLMs tend to add explanations and markdown around JSON responses. Solved this by being extremely explicit in system prompts: "Return ONLY valid JSON. No explanation. No markdown backticks."

Graph coloring timing — making sure the knowledge graph node colors updated after the misconception detector returned scores, not before, required careful async state management.

Balancing question depth vs session length — the Socratic Examiner would sometimes go too deep on one concept. Added a next_concept_to_test field in the detector's JSON response to guide the flow forward.

What I Learned

Multi-agent prompt engineering is an art — each agent needs a very tight, specific system prompt or outputs drift
The confidence vs competence gap is real and measurable even in a short 7-exchange session
Lovable + Groq is a surprisingly powerful stack for rapid AI app prototyping