Focus-AI

Adaptive learning that detects confusion and responds in real time.

---

Focus-AI is an adaptive tutoring platform that uses behavioural telemetry (hesitation, scrolling, tab switches) to detect when a student is struggling, then a contextual bandit selects the best explanation strategy. No cameras, no wearables — just the signals a browser already provides.

What it does

• Detects confusion in real time — XGBoost classifier on interaction signals, inference in under 12ms
• Adapts content dynamically — 5 difficulty levels per topic, switched instantly by a LinUCB bandit
• Voice-first accessibility — Navigate, quiz, and ask questions entirely by speech
• AI tutoring — Free-form questions answered by AI grounded in the current lesson
• 12 free LLMs — Cost-aware routing through OpenRouter; 96% cheaper than all-paid
• Deep Knowledge Tracing — Per-skill mastery tracked over time with LSTM-based BKT
• 6 CS courses — DBMS, DSA, OS, Computer Networks, OOP, Software Engineering

Tech stack

Frontend	Next.js 16, React 19, TypeScript, Tailwind CSS v4
Backend	FastAPI, Python 3.10+, Pydantic v2
Database	Firebase Firestore
Auth	Firebase Auth
ML	XGBoost, PyTorch (LSTM), scikit-learn, custom LinUCB
AI	Gemini 2.5 Flash + OpenRouter (12 free models)
UI	Radix UI, Framer Motion, Lucide Icons

Quick start

# Clone
git clone https://github.com/MadeNavaneeth/Focus-AI.git
cd Focus-AI

# Backend
cd backend
python -m venv .venv && .venv\Scripts\activate
pip install -r requirements.txt
cp .env.example .env          # Fill in your Firebase + API keys
uvicorn main:app --reload --port 8000

# Frontend (new terminal)
cd frontend
npm install
# Create frontend/.env with your Firebase web config
npm run dev

Open http://localhost:3000. Sign up, and the adaptive loop starts working immediately.

How the adaptation loop works

Student interacts with lesson
  → useBehavior hook captures dwell time, scrolls, tab switches
  → POST /api/v1/behavior
  → XGBoost predicts confusion probability
  → If confused: LinUCB bandit selects best explanation arm
  → Content level shifts instantly (Expert → Simplified → Analogy)
  → Reward signal updates the bandit for next time

All under 200ms end-to-end.

Voice commands

Press Left Ctrl or click the mic button:

Say	Action
"Open DBMS"	Navigate to a specific course
"Take the quiz"	Start the lesson quiz
"Option B" / "Submit"	Answer quiz questions
"Read this lesson"	TTS narration
"Explain normalization"	AI tutor answers from lesson context
"Stop"	Cancel speech

Environment variables

Backend (backend/.env):

FIREBASE_ADMIN_CREDENTIALS=firebase-adminsdk.json
GEMINI_API_KEY=your-key          # Optional (paid fallback)
OPENROUTER_API_KEY=your-key      # Optional (free models)

Frontend (frontend/.env):

NEXT_PUBLIC_API_URL=http://localhost:8000/api/v1
NEXT_PUBLIC_FIREBASE_API_KEY=your-firebase-web-key
NEXT_PUBLIC_FIREBASE_AUTH_DOMAIN=your-project.firebaseapp.com
NEXT_PUBLIC_FIREBASE_PROJECT_ID=your-project-id

Project structure

Focus-AI/
├── backend/          FastAPI REST API + ML inference
├── frontend/         Next.js 16 web app
├── ml/               Training scripts + serialized models
├── scripts/          Data seeding + visualization
└── seed/             Initial course content (JSON)

Authors

• Yadamreddy Navaneeth
• Shoham Mohapatra
• Ekaksh Agrawal
• Dr. Eleena Mohapatra (Faculty Advisor)
• Dr. Pratiba K (Faculty Advisor)

License

MIT