EduTrack

Inspiration

11% of U.S. kids have ADHD, and even neurotypical students lose attention every 4–5 minutes of a lecture. Classrooms still teach one lesson, one pace, to thirty different brains — and we don't see why a student struggles until the test is graded. We wanted to flip that.

## What it does

EduTrack adapts lessons in real time using how students actually read.

Students read while a tracker streams mouse, scroll, hover, idle, and re-read events. Each session is classified into 5 focus states and feeds a 7-axis learning profile + topic-mastery graph.
Educators open Lesson Studio: type a topic → AI generates reading variants, quizzes, practice, and YouTube videos. Drag, reorder, publish.
Researchers get a per-student Neural Microscope (44→10→6→1 net with full saliency + heatmaps) and a 3D TRIBE v2 brain that predicts which ROIs activate per lesson.

## How we built it

Backend: FastAPI + async SQLAlchemy + pgvector + Redis + Celery + XGBoost.
Frontend: Next.js 14, React 18, Three.js for the brain, Recharts.
Adaptation: continual and persistent learning through a blend of RAG and RL — every session updates the style vector and mastery graph; retrieval feeds the next AI-generated lesson via Claude.
Per-student NN: pure NumPy SGD, student samples weighted 3× cohort, fully explainable.
Tests: pytest + Vitest + Playwright across unit, integration, and E2E.

## Challenges we ran into

Threading session_id from lesson page through quiz to close the prediction loop.
Cold-start fallbacks everywhere — deterministic embeddings, heuristic backprop targets, simulated BOLD — so the demo never goes blank.
Running async coroutines from sync Celery contexts without dropping WebSocket events.
Making a neural network readable for non-ML researchers.
Keeping pgvector and SQLite RAG parity so verification works without Docker.

## Accomplishments that we're proud of

A real adaptive loop — behavior → prediction → quiz → profile → personalization — every step inspectable.
A 3D brain grounded in anatomically-placed ROIs, tied to lesson content.
An explainable per-student NN with live saliency and weight heatmaps.
Topic → publishable lesson plan in under 2 minutes.
Full test coverage across unit, API, worker, component, and E2E layers.

## What we learned

Behavior data predicts comprehension better than scores — and lets us intervene during learning, not after.
Adaptive systems show ~0.40 effect size on achievement; for executive-function challenges, d ≈ 0.76 (50th → 77th percentile).
Explainability builds trust. Teachers adopt personalization when they can see why.
Graceful degradation matters as much as the happy path.

## What's next for EduTrack