math-genome-ai

Stop guessing why you're stuck. Math Genome AI diagnoses, adapts, and guides you to your target score with persistent memory.

Comment

Inspiration Most math apps mark answers right or wrong. We watched students repeat the same mistake — not because they don't know algebra, but because they mis-translate word problems, or make sign errors under time pressure. We wanted an AI that diagnoses why, not just what.

That's Math Genome AI — a true autonomous learning agent that builds a persistent "learning DNA" for every student.

What it does Math Genome is NOT a chatbot. It's a closed-loop agent that:

  1. Math DNA Scanner – 7-dimension diagnostic (Concept Understanding, Calculation Accuracy, Pattern Recognition, Logical Reasoning, Word Problem Translation, Time Efficiency, Confidence)

  2. Root Cause Engine – For every mistake, it doesn't say "weak in algebra." It says:

Primary Cause: Equation Translation (91% confidence) Secondary: Reading Comprehension

  1. Autonomous Recovery Agent – Decides next action with reasoning:

ASK_EASIER_QUESTION, PROVIDE_HINT_MODE, Recovery Session, etc. Always explains WHY

  1. Learning DNA Map – Live visualization of strengths, weaknesses, and hidden patterns

  2. Future Performance Predictor – Projects scores: 7-day, 14-day, 30-day using genome trends

  3. Failure Recovery Mode – After 3 consecutive fails, auto-switches teaching strategy (Visual → Step-by-Step → Scaffold)

  4. Math GPS & Daily Missions – Personalized path from 540 → 750 with gamified daily tasks

The agent loop runs after EVERY interaction: Answer → Evaluate → Update Genome → Root Cause → Decide Action → Next Step

How we built it Frontend: Next.js 14, React, Tailwind CSS, Recharts for genome radar Backend: FastAPI (Python) – agent brain AI Layer: LLM for root-cause analysis, question generation, and recovery planning Database: MongoDB for persistent Math Genome profiles Deployment: Bolt.new → Vercel (frontend), Railway (backend)

Architecture:

Student State Memory stores 7 genome dimensions + mistake history Decision Engine prioritizes weakest dimension always Skill update: correct = +5 to +15, wrong = -5 to -10 Challenges we ran into The biggest was Next.js build crash – Radix UI Progress component's minified CJS build contained unescaped backticks that crashed Node's parser during page-data collection. We fixed it by reimplementing Progress as a lightweight div-based component.

Second challenge: making root-cause diagnosis accurate. We built a rule-based engine first (sign error, fraction manipulation, time pressure) then layered LLM reasoning on top for 90%+ confidence.

Third: keeping the agent truly autonomous. We had to prevent it from becoming a Q&A bot by enforcing the decision loop and "why_this_action_was_taken" field in every API response.

Accomplishments that we're proud of Agent makes 100% of learning decisions without human input Persistent genome that evolves across sessions Real-time strategy switching after failures Built in 36 hours on Bolt + FastAPI What's next LLM-generated infinite question bank Multiplayer streaks and classroom dashboards Voice-guided tutor mode Integration with SAT official practice tests

Built With

Share this project:

Updates