ClassBridge

Inspiration

Every student deserves to understand what they're learning — but millions of kids in India and around the world receive worksheets written in English, at grade levels above their current understanding, with no support in their home language. Parents can't help. Teachers are stretched thin.

We built ClassBridge to close that gap.

What it does

ClassBridge takes any school worksheet — pasted text, an uploaded file, or a built-in demo — and transforms it into a fully personalized learning pack in seconds:

• Simplified Explanation — broken down to the student's exact grade level with real-life analogies
• Bilingual Translation — side-by-side English + Tamil/Hindi so nothing is lost
• Study Cards — bite-sized flashcards categorized as definitions, concepts, formulas, or examples
• Practice Questions — multiple choice, short answer, and open-ended with instant AI feedback
• Parent Summary — key concepts, difficult areas, and revision tips written for parents and teachers ## How we built it The core loop is a single /api/generate-learning-pack endpoint. It accepts the worksheet text plus the student's grade level, language, and learning mode, then sends a carefully engineered prompt to the NIM API. The response is validated against a strict Zod schema before any data reaches the UI — ensuring the AI never returns malformed content silently. ## Challenges we ran into Structured LLM output at scale. Getting a language model to return seven nested JSON sections reliably — explanation, translation, study cards, practice questions, and more — without hallucinating field names or breaking schema required careful prompt engineering and Zod validation with graceful error handling.

Bilingual generation quality. Tamil and Hindi educational translations needed to be natural, not robotic. We tested multiple prompt framings before settling on one that instructs the model to translate meaning, not words.

Practice question grading. Normalizing student answers for short-answer questions (case, punctuation, partial matches) without a full NLP pipeline meant building a clean normalizeAnswer utility that handles the most common edge cases gracefully.

What we learned

How to extract reliable structured data from LLMs using Zod schemas as a contract

• NVIDIA NIM API integration and prompt engineering for educational content
• Building accessible, multilingual UIs with a focus on real student needs
• The importance of graceful degradation — every AI failure has a user-facing fallback ## What's next for ClassBridge OCR pipeline for uploaded PDFs and images (currently paste/demo only)
• Voice read-aloud in Tamil and Hindi using TTS
• Teacher dashboard for class-wide worksheet uploads
• Offline-capable PWA for low-bandwidth school environments

Built With

• nextjs
• node.js
• nvidia
• react
• tailwindcss
• typescript
• vercel
• zod