Inspiration
It started with the classmates I sit next to in high school. Many of them struggled with high school math, and when I sat down to help them, I noticed something: their problems rarely began with the topic they were stuck on. When we traced their mistakes backwards, the real gaps were much earlier: basic algebra, factoring, and even basic integer operations.
A wrong answer in a Limits problem is often not a Limits problem at all. And students usually don't know which earlier skill is missing. I built MathMap so they can find out, before they run into advanced topics like Limits.
What it does
MathMap AI is a math learning app that finds the root cause behind a student's mistakes, not just whether an answer is right or wrong.
- Diagnose: a short test where every wrong option maps to one specific misconception.
- Map: mistakes are traced through a prerequisite graph (Integer Operations → Basic Algebra → Factoring → Quadratic Equations, with Functions and Limits building on them), producing a learning path with a clear "Start here".
- Learn and practice: short lessons, five fresh practice questions weighted toward the student's own mistakes, and a per-question AI tutor. Mastery changes (for example Needs work → Developing) are shown after practice.
- Navigator: a planning agent. A student can say "I only have 10 minutes today" or "Am I ready for Limits?", and Navigator chooses tools, reads the student's data, and builds a plan around their study time. The UI shows which tools it actually called and which model answered.
How we built it
Principle: the deterministic core is the source of truth. The question bank, scoring and prerequisite graph decide what is correct, what is mastered and where a mistake comes from. Language models only provide language and planning. A model never decides right or wrong.
- Backend: FastAPI with SQLite. Frontend: vanilla HTML, CSS and JavaScript, with KaTeX for math.
- Everything runs on Nebius Token Factory.
- NVIDIA Nemotron 3 Super runs Navigator, a tool-calling agent with six read-only tools. It also blind-verifies every generated question: it solves the question without seeing the answer key and must agree that exactly one option is correct.
- DeepSeek V4 Pro writes lessons, generates questions and powers the per-question tutor. We chose it for the tutor after comparing it directly with Nemotron on the same questions, because its diagnosis of the student's mistake was more accurate.
- The agent is read-only. It can only propose actions as buttons, and the student decides.
- Deployed on Render, with the whole project built from an Android phone using Replit and the GitHub web editor.
Challenges we ran into
- Deciding what the AI should and should not do. Early versions let the model judge answers, and it was inconsistent. Moving scoring, the graph and misconception codes into deterministic code, and leaving the model only the language and planning, made the app trustworthy.
- Question quality. LLM-generated questions can be subtly broken. Structural validation plus blind verification by Nemotron caught the bad ones, and in manual review every question the verifier flagged was genuinely broken.
- Messy model output. Models sometimes return math in different LaTeX formats, so the client normalizes it before rendering.
- Building and hosting from a phone. Pasting long scripts into a mobile shell often got truncated, so we moved to small, checked patches. Several free hosting options failed before we got a working deployment on Render's free tier.
Accomplishments that we're proud of
- A working app, publicly deployed and open source (MIT), that anyone can open without logging in.
- An agent that visibly uses tools instead of just chatting, with a transparent tool trace and model footer in the UI.
- A diagnosis flow that connects every mistake to a specific misconception and a prerequisite skill.
- Building the whole thing as a first agentic AI project, entirely from a phone.
What we learned
A lot, since this is my first agentic AI project, and I don't come from a professional engineering background. I'm just a high school student who is curious about AI and how it can solve real problems for people.
I learned how an agentic AI workflow fits together: how to structure the logic, and how to make every decision serve the original goal without drifting off track. I also picked up many new tools, from getting comfortable with GitHub and building in Replit to deploying on Render. And I learned to compare AI models and choose the right one for each task in the project, instead of using one model for everything.
What's next for MathMap AI
- Dedicated topic question banks for every concept (today Limits and Basic Algebra have one).
- A persistent database so progress survives server restarts. On the free demo host, progress can reset when the server restarts.
- Reviewing the remaining lessons, with a Nemotron check.
- Navigator in more languages, starting with Indonesian.
- Optional accounts, so progress follows a student across devices.
Built With
- deepseek
- fastapi
- html
- javascript
- nebius
- nemotron
- nvidia
- python
- render
- replit
- sqlite
