ClassBox

Teacher session creator
Teacher view - detailed student view
Teacher view - no stats
Landing page
Teacher view - with stats
Project-friendly session link
Student view

Inspiration

Students can’t avoid AI.

In today’s classrooms, teachers are forced into a losing game: either ban AI entirely or spend time trying to catch students using it. AI detection tools are unreliable, reactive, and miss the point - students will use AI regardless.

We realized the real problem isn’t AI usage, it’s how students use it.

What if, instead of fighting AI, we taught students how to use it as a learning assistant?
And what if teachers could finally see how students think, not just what answers they submit?

That insight led to ClassBox.

What it does

ClassBox is a real-time AI tutoring platform designed for classrooms and assessments, built around the Socratic method and full teacher transparency.

Instead of giving direct answers, ClassBox:

Guides students through problems using AI-generated questions
Encourages reasoning and conceptual understanding
Prevents shortcut-seeking behavior with built-in guardrails

At the same time, teachers get:

A live dashboard showing student activity and engagement
Visibility into how students are interacting with AI
Control to pause AI, upload course materials, and enforce strict context usage
Aggregated insights into where students are struggling

By making ClassBox the official AI tool used during assessments, students are incentivized to also use it at home, creating consistent learning behavior and meaningful insights for educators.

How we built it

ClassBox is a full-stack, real-time application built with production-ready architecture:

Frontend: React + Vite + Tailwind CSS
Backend: Node.js, Express, and Socket.io for real-time communication
Database: MongoDB Atlas with TTL indexes for auto-expiring sessions
AI: GPT-4o for Socratic tutoring
RAG: OpenAI Vector Stores for document-based context retrieval
Context Compression: The Token Company API to reduce token usage and costs

The AI pipeline follows five stages:

Session pause checks
Input validation and guardrails
Context compression
Retrieval-augmented generation using teacher-uploaded documents
Socratic response generation with streaming output

All student interactions are stored directly within MongoDB documents for fast retrieval, live dashboards, and accurate metrics.

Challenges we ran into

Preventing direct answers: Designing AI prompts and guardrails that reliably guide students without leaking solutions required multiple iterations.
Making the AI sound natural: It was harder than one would expect to create a chatbot that takes various contexts into account while also respecting guardrails and not be frustratingly repetitive all at the same time
Synchronization & Networking: Keeping teacher dashboards, student chats, and AI streams perfectly in sync using WebSockets was technically challenging.
Balancing transparency and privacy: We wanted teachers to gain insight without creating a surveillance tool or penalizing individual students.

Accomplishments that we're proud of

Building a fully operational, real-time AI tutoring platform in 24-hours
Implementing a RAG pipeline with teacher-uploaded documents
Designing AI guardrails that promote learning instead of policing behavior
Creating meaningful teacher-facing insights without increasing workload
Attempting to deliver a production-ready architecture, not just a demo

What we learned

AI in education isn’t a technical problem, it’s a pedagogical design problem
Students respond better to guidance than restriction
Teachers don’t need more grading tools, they need better insight
Real-time systems require thoughtful data modeling and failure handling
Guardrails are more effective than detection when shaping behavior
Hackathons are a great place to build your MVPs