Inspiration
Every classroom has students who struggle silently. Some hesitate to ask questions because they're afraid of being judged. Others understand the subject but struggle because it's being taught in a language they're less comfortable with. Meanwhile, teachers often have no way of knowing which concepts their students truly misunderstand until it's too late.
We wanted to answer a simple question: What if every student could learn in their preferred language, while teachers could understand what their entire classroom was struggling with?
That's what inspired ClassBridge.
What it does
ClassBridge is an AI-powered inclusive education platform that connects three portals — teacher, student and parent — around one idea: every student should be able to ask, and every teacher should be able to understand.
It tackles three learning barriers: language differences, fear of asking questions, and hidden conceptual misunderstandings.
🏫 Institution accounts and courses
Teachers, students and parents sign in with accounts issued by their school or university — no self sign-up, no join codes. Students see the courses they're enrolled in and join a live lesson from the course page in one click. Each course has its own live lessons, homework, quizzes and released grades.
🌍 Multilingual learning
Students choose from 30+ languages and can switch at any moment without losing their questions or progress. Lesson notes and AI answers are translated on demand and cached, so switching back is instant.
They can ask private questions in any language or script — including romanized languages like Hinglish — regardless of their interface language. Questions are stored exactly as typed.
Teachers see every question translated into English, with the original one click away. If a translation fails, the original is still shown with a clear "translation unavailable" state and a retry button.
🎯 Cross-Language Misconception Radar
Our most distinctive feature. Instead of translating or answering questions one by one, the AI reads all of a session's questions together with the lesson notes, and groups them by the misunderstanding behind them — not by wording or language.
For example, students might ask:
- English: "Why is 1/8 smaller than 1/4?"
- Hindi: "बड़ा हर होने पर भिन्न छोटी क्यों होती है?"
- French: "Pourquoi un plus grand dénominateur donne-t-il une plus petite fraction ?"
- Hinglish: "Agar denominator bada hai toh fraction chota kyun hota hai?"
ClassBridge recognises one shared misconception — a larger denominator means a larger fraction — links all four questions as evidence, and suggests an intervention such as a visual fraction comparison. "How do I multiply two fractions?" stays separate, even though it's also about fractions, and off-topic questions are left unclustered.
The radar updates during the live lesson as new questions arrive. Teachers can inspect exactly which questions support each misconception, and every result is labelled as an AI suggestion — the teacher decides what to do.
🧩 Teacher-approved mini-lessons
From a detected misconception, the teacher can generate an interactive mini-lesson with an explanation in each student's language, plus a short before and after understanding check. The teacher previews, approves and shares it. Mini-lessons run in a sandboxed frame with no network access, and the dashboard shows the real before/after results — never invented numbers.
✋ Live understanding check-ins
During a lesson, students can privately signal whether they understand, need an example, or are confused, and the teacher sees a live summary next to the question feed.
📚 Homework, quizzes and grades
Teachers create homework and quizzes for each course. Students submit and take them from their course page; multiple-choice questions are marked automatically on the server, using an answer key students can never read. Teachers grade, add feedback and choose when to release grades — only released grades are visible to students and parents. Results also feed a performance view that highlights topics needing attention.
👪 Multilingual parent portal
Parents see released grades, homework and areas where their child needs support, in their own language — and only for children their institution has linked to them.
💡 Personalized explanations
Students can ask for a simpler explanation or a concrete example of the current lesson, in their chosen language, grounded in the teacher's lesson notes.
How we built it
We designed the user experience in Lovable, then exported the project to GitHub and continued development with Cursor and Claude.
- Frontend: React, TypeScript, TanStack Router, Tailwind CSS and shadcn/ui
- Backend: Supabase — PostgreSQL with row-level security, Supabase Auth for institution accounts, and Edge Functions (Deno) for every AI call, so API keys never reach the browser
- AI: an OpenAI-compatible model via Groq, with every response validated against a Zod schema before it's stored
Our question pipeline
- Save first — the question is stored exactly as typed, before any AI call, so nothing is lost if the AI fails.
- Detect & translate — the AI identifies the language (including romanized scripts) and writes a faithful English translation that keeps the student's meaning, uncertainty and maths notation.
- Answer — a separate call answers the student in their chosen language, grounded in the lesson notes.
- Analyse the session — one call reads all questions with the lesson context and returns misconception clusters as validated JSON. A misconception needs at least two supporting questions; unrelated or off-topic questions stay unclustered. Analysis is debounced to avoid unnecessary calls.
- Teacher in the loop — the teacher reviews the evidence, previews a mini-lesson and decides what to share.
Access is enforced in the database: courses, rosters and parent links come from an institution directory; students only see their own courses and work; teacher-only actions are verified on the server; and quiz answer keys and unreleased grades are hidden from students and parents.
Challenges we ran into
- Never losing a question. Our first version translated before saving, so an AI error meant a lost question. We redesigned the pipeline to store first, track translation status, and let teachers retry.
- Same topic vs. same misconception. Two questions about fractions can reveal completely different misunderstandings. Keyword grouping would give useless insights, so we analyse with the lesson notes, require evidence from at least two questions, and tested that unrelated questions stay apart.
- Romanized languages. Hinglish looks like English to simple detectors — we had to make detection and translation treat it as Hindi written in Latin letters.
- Trust and privacy. Without a full school identity system, we still needed students to see only their own work, parents only released grades, and quiz answers to stay hidden — solved with row-level security and server-side functions.
- Time and scope. With two developers and a hackathon deadline, we focused on making the AI pipeline genuinely work end to end, with clear loading and error states, instead of building every possible feature.
Accomplishments that we're proud of
In live end-to-end tests, questions in English, Hindi, French, Spanish and Hinglish — including phrasings that never appeared in our prompts — were translated faithfully and grouped under the correct shared misconception, while the unrelated multiplication question stayed separate. Every AI result in the app comes from a real model call, not hardcoded answers.
What we learned
- The value of educational AI isn't just generating explanations — it's turning student input into insights teachers can act on.
- Designing for failure (save first, show clear error states) matters as much as the AI itself.
- Teachers trust AI more when it shows its evidence and leaves the decision to them.
- We gained hands-on experience with AI-assisted development, connecting a Lovable frontend to a real backend, and designing multilingual interfaces.
- Most importantly: accessible education isn't just translating content. It's creating an environment where every student can communicate, participate, and be understood.
What's next for ClassBridge
- AI-assisted grading and personalized feedback on homework
- More interactive lesson templates beyond fractions
- Instant live updates with Supabase Realtime (currently short polling), and syncing understanding check-ins across devices
- School single sign-on and admin tools for managing rosters
- Cloud storage for lesson-plan and homework PDFs
- Long-term progress tracking that shows whether an intervention actually worked
Why ClassBridge matters
AI in education shouldn't just make answering questions faster — it should help educators understand their students better. ClassBridge lets students learn and ask in their own language, while giving teachers real-time insight into what their class finds difficult.
Because every student deserves to understand, not just attend.
Built With
- chatgpt
- claude
- cursor
- deno
- groq
- llm
- lovable
- postgresql
- react
- shadcn-ui
- supabase
- tailwindcss
- tanstack-router
- typescript
- vite
- zod
