Noteforge-AI

🎯 Problem Statement

Students, researchers, and professionals often spend hours manually reading large documents, preparing notes, creating quizzes, extracting important information, and converting content into different learning formats. Existing AI tools usually provide only one type of output or require users to move between multiple platforms. This makes the document-learning process slow, repetitive, and difficult to manage. NoteForge was created to provide one central platform where users can upload source material and transform it into useful, source-grounded learning content.

🛠️ What I Built

I built NoteForge, an AI-powered knowledge and document-learning platform that connects a custom web portal with Google NotebookLM through browser automation and backend services. Users can upload PDF, DOCX, TXT, Markdown, CSV, and image files or provide a webpage URL. They can then choose from multiple output formats, including: Summaries Study notes Questions and answers Briefing documents Quizzes Flashcards Timelines Mind maps Infographics Slide decks Audio overviews Video scripts Data tables Custom prompts The system creates a NotebookLM notebook, uploads the selected source, waits for it to be processed, requests the selected output, and returns the generated result to the NoteForge interface. I also developed: A public landing page A document upload interface Output-category navigation Notebook and source management Task-processing status updates Generation history Result preview and copying Download options for supported outputs Cookie-based authenticated NotebookLM sessions Backend health and connection checks A local runner and server-based task workflow Error handling for failed uploads and generation tasks The platform uses Playwright automation, Python backend services, Node.js components, and the notebooklm-py library to communicate with NotebookLM workflows.

🔥 Challenges I Faced

The most difficult part was creating a reliable connection between the custom NoteForge interface and NotebookLM. Google restricts automated logins from VPS and datacenter IP addresses, so normal browser automation could not consistently authenticate. I experimented with persistent browser profiles, VNC sessions, headless browsers, cookie imports, and a local runner architecture. Another major issue was automating file uploads. NotebookLM's interface and selectors changed, and sometimes the file chooser did not appear after the upload button was clicked. I had to improve selector handling, session validation, upload detection, retry logic, and error reporting. Audio, infographic, quiz, flashcard, and other Studio outputs also followed different generation workflows. Some were generated directly, while others required navigating through the NotebookLM Studio interface and waiting for asynchronous processing.

💡 What I Learned

This project taught me that building an AI product involves much more than creating a user interface. I gained practical experience with: Browser automation using Playwright, using hermes, and used a notebooklm python library for improving reponse time I also learned the importance of validating the complete workflow rather than testing only individual buttons. A feature is only complete when its frontend, API, browser automation, generated output, error handling, and download process all work together.

🚀 Future Improvements

Have no upload limits and Cloud file storage

🧰 Tech Stack & AI Tools

🤖 AI Reviewer Feedback

You built an impressive end‑to‑end platform that integrates many output formats and tackles the tricky automation challenges of NotebookLM, showing solid full‑stack and DevOps skills. To strengthen the product, consider adding automated tests for the Playwright flows and a more robust authentication method to avoid reliance on cookie hacks, which will improve reliability and scalability. Overall, great work with clear problem focus and valuable learnings.