Offline AI education
1st place, EduThon, IIT Bombay TechFest (29th edition)
MasterG
Curriculum-aligned lessons generated in 22+ Indian languages, fully offline, on hardware that has never heard of a GPU.
Problem
Most AI tutoring assumes a fast connection, a strong device, and English. The classrooms that would benefit most have none of the three. Content also has to follow the actual curriculum, not whatever the model feels like teaching.
MasterG had to generate curriculum-aligned material in the languages students actually speak, with no internet at all, on low-resource machines.
Approach
Everything runs locally. Ollama serves the language model on the device itself, ChromaDB holds the curriculum as a local vector store so generation stays aligned to the syllabus, and MongoDB stores the produced content. There is no cloud call anywhere in the path.
The generation layer targets 22+ Indian languages, and the React and TypeScript front end is kept deliberately light so it stays responsive on the kind of hardware schools already own.
Architecture
1. Run locally
The whole stack lives on the device. Ollama serves the model with no network in the loop.
2. Ground in the syllabus
ChromaDB holds the curriculum as a local vector store so every lesson stays aligned to what is actually taught.
3. Generate in 22+ languages
Content is produced in the student's language and stored in MongoDB.
4. Serve
A light Node.js API and a React front end keep it usable on modest hardware.
Stack
- React
- TypeScript
- Node.js
- Ollama
- ChromaDB
- MongoDB
Outcome
A working, fully offline educational platform that generates curriculum-aligned content in 22+ Indian languages. It won first place at EduThon, IIT Bombay TechFest's 29th edition.
What I'd do differently
Offline-first is a constraint that improves the product. Every shortcut that needed the cloud was off the table, so the retrieval layer had to be good enough to carry generation quality on its own. It was.