Astra - Won Best Use of MongoDB @ HackAI 2026

Immersive 3D UTD learning galaxy: course → concept graph, Bayesian mastery, per-node quizzes, and explainable recommendations.

Stack

• Frontend: Vite + React + TypeScript, ForceGraph3D, Zustand, Framer Motion
• Backend: FastAPI, Gemini (quiz/concept extraction), NV-Embed-v2 → 1024-dim projection, Beta-Binomial mastery
• Data: Demo data (public/demo_data.json), optional Nebula API, RMP pre-scraped JSON

How to run

1. Frontend (required)

cd nebula-galaxy
npm install
npm run dev

Open http://localhost:5173. Complete onboarding, then explore the galaxy. Click a node for details and “Generate quiz”; for courses, use “Open concept map” for the Level 2 concept map.

2. Backend (optional – for real quizzes and mastery API)

In a second terminal:

cd nebula-galaxy/backend
python -m venv venv
source venv/bin/activate          # Windows:  venv\Scripts\activate
pip install -r requirements.txt

Add your API keys in backend/.env (copy from backend/.env.example if needed):

• GEMINI_API_KEY – for quiz and concept generation (Google AI Studio)
• NEMOTRON_API_KEY – for NVIDIA NIM (e.g. NV-Embed-v2 embeddings at NVIDIA Build)

Then start the API:

uvicorn main:app --reload --port 8000

The frontend proxies /api to http://localhost:8000. Without the backend, the app still runs with demo data and local Bayesian fallback.

Data scripts (run before demo)

Nebula API – courses, professors, grade distributions:

pip install httpx
python scripts/fetch_nebula.py --output public/nebula_data.json

Fetches /course/all, /professor/all, and per-course /course/{id}/grades and /course/{id}/professors. No API key required for the public Nebula API.

Rate My Professor – scrape UTD professor ratings (clarity/difficulty) for insights:

python scripts/fetch_rmp.py --output public/rmp_data.json

If scraping fails (RMP may block or change structure), the script writes a template JSON; you can add entries manually. The app and backend can load rmp_data.json and nebula_data.json for course details and for Gemini when the user uploads a syllabus (Gemini extracts concepts and can use this metadata for recommendations).

Environment

• Root .env – frontend: VITE_NEBULA_API_KEY, VITE_API_URL (defaults work with proxy to backend).
• backend/.env – backend: GEMINI_API_KEY, NEMOTRON_API_KEY, NVIDIA_NIM_BASE_URL, MONGODB_URI. Embedding dimension is locked to 1024; NV-Embed-v2 4096-dim vectors are projected to 1024 before storage.

Features

• Level 1 – Galactic degree plan: UTD courses, prereq edges, node glow from mastery
• Level 2 – Planetary concept map: Concepts per course, centripetal force so concepts stay near the course
• Dual-glow: UTD average success vs your personalized mastery (Beta-Binomial with slip/guess)
• Per-node panel: Quiz generation (Gemini), mastery update on answer, “Open concept map” for courses
• Breadcrumbs & Back/Home: Galaxy → Course → Concept; warm start when returning to Level 1
• Concept pipeline: Syllabus PDF or Nebula description → Gemini extracts 5–8 concepts → prefixed IDs (concept_<course>_<slug>)

Project layout

nebula-galaxy/
├── src/
│   ├── components/   # GalaxyGraph, ConceptMap, NodePanel, Breadcrumbs, OnboardingForm
│   ├── store/       # Zustand (graph, mastery, breadcrumbs, positions)
│   ├── data/        # loadGraph, demo data loader
│   ├── lib/         # api, bayesian, constants (EMBED_DIM=1024)
│   └── types/       # graph, user, api
├── public/
│   └── demo_data.json
├── backend/
│   ├── main.py
│   ├── config.py    # EMBED_DIM, NV_EMBED_RAW_DIM
│   ├── routers/    # generate-node-content, predict, bayesian-update, nebula/courses
│   └── services/   # gemini_service, bayesian_service, embedding (PCA 4096→1024)
└── README.md

Hackathon tracks

• Nebula Labs: Course/concept decisions, DFW, scheduling, nested course → concept visualization
• Dallas AI: Prompt anchors (tone by mastery), feature attribution, explainable next steps
• ML (MLH): Gemini (quizzes, concept extraction), NV-Embed-v2 (embeddings, 1024-dim), Beta-Binomial mastery with slip/guess