NoteForge: Offline-First Intelligent Learning Platform

#mlhacks #devchallenge #hackathon #gemma

This is a submission for the MLH x DEV Writing Challenge

NoteForge — Transform Notes. Forge Knowledge.

What I Built

Modern students and lifelong learners face an ongoing paradox: lecture decks, academic papers, and textbooks have never been more accessible, yet turning massive PDF dumps into active, structured knowledge remains tedious and friction-heavy. Most existing AI note-converters require persistent cloud connections, risk private student data leakage, or operate as basic wrappers around remote LLM APIs.

I built NoteForge: a production-grade, offline-first intelligent learning platform designed to transform uploaded lecture notes, educational slide decks, and research PDFs into an interactive learning ecosystem without sending a single byte of your data to external servers.

Key Capabilities:

Offline-First Grounded AI Engine: Powered by locally hosted Gemma 2 running through Ollama. Every generated concept, definition, and question is strictly grounded in the uploaded material with zero hallucination.
3D Interactive Flashcard Suite: Immersive 3D flip card interactions with keyboard navigation (Space to flip, K to mark mastered, R for review), deck shuffling, bookmarking, and Leitner-style spaced repetition scheduling.
Quiz Arena: Timed assessments with multiple-choice, true/false, and conceptual questions, instant scoring, accuracy breakdown, and question-by-question reviews citing direct explanations from your notes.
Smart Study Notes: Automated synthesis of dense material into executive summaries, key takeaway matrices, formal definitions, and high-yield examination tips with one-click Markdown export.
Learning Analytics Engine: Tracks study duration, streak habits, flashcard retention distribution, and diagnostic concept error analysis derived from real study sessions.
Data Sovereignty: Full JSON workspace backup and restore, local SQLite database storage, and complete offline capability.

Demo

GitHub Repository: https://github.com/Babin123456/NoteForge

Highlights & Workflow:

Document Library: Drop in lecture PDFs of any size. PyMuPDF extracts and normalizes the structural text locally in milliseconds.
Curriculum Synthesis Matrix: Choose your study goal (Flashcard Deck, Timed Quiz, or Smart Notes), adjust target size and difficulty, and let local Gemma 2 synthesize tailored study tools.
Spaced Repetition & Arena: Test comprehension with instantaneous feedback and track your weekly momentum on the interactive dashboard.

Partner Technologies

1. Gemma 2 (Google DeepMind)

NoteForge uses Google's lightweight open model family—specifically Gemma 2 (via Ollama)—as its core intelligence engine.

Local Native Execution: Running Gemma 2 locally ensures NoteForge works seamlessly on flights, in lecture halls with unstable Wi-Fi, or in offline study sessions.
High-Fidelity Structured Output: By enforcing structured JSON schemas in Gemma 2's prompt templates, NoteForge extracts complex taxonomies, formula relationships, and distractor choices for quizzes with zero external API latency or subscription fees.
Local Fallback Grounding: Coupled with a deterministic local extraction engine, NoteForge guarantees uninterrupted study deck creation even if models are reloading.

2. Modern Full-Stack Tooling

FastAPI & PyMuPDF: High-performance asynchronous backend pipeline for low-latency PDF parsing and SQLite persistence.
React 19 + TypeScript + Vite + Tailwind CSS: Fast, responsive, dark-first UI featuring glassmorphic components, Framer Motion transitions, and Recharts data visualizations.

Hackathon Experience

Building NoteForge during the hackathon was an exhilarating deep dive into offline-first edge AI architecture.

The biggest challenge was bridging the gap between local LLM constraints and high-speed UI responsiveness. Optimizing PyMuPDF text chunking, designing strict JSON output validation for Gemma 2, and ensuring smooth 60fps 3D card flips taught me a lot about responsive UI design and client-side state synchronization.

What I'll remember most is the realization that developers don't need to depend on proprietary cloud APIs to deliver an extraordinary, private, and lightning-fast learning tool. Building an offline learning cockpit that puts complete data ownership back in students' hands was the most rewarding part of the hackathon!

DEV Community