AI-Powered Quiz Generator
An automated quiz system using LangGraph agents. It generates questions, checks if they're any good, and keeps refining them until they meet quality standards.
How It Works
Built a LangGraph agent that generates quiz questions from input material, scores them on clarity and relevance, then re-prompts with feedback if they're not good enough. This feedback loop produces way better questions than just asking once and calling it done.
Technical Details
LangGraph Agent Architecture
The agent runs through multiple nodes: 1. Generation Node: Creates quiz questions from the input 2. Evaluation Node: Scores questions on clarity, difficulty, and relevance 3. Refinement Node: Figures out what's wrong with bad questions and generates improvement suggestions 4. Decision Node: Checks if we've hit the quality threshold or need another pass
LangGraph's state management made this way easier to build than I expected. Before using it, I thought managing agent workflows would be a mess.
OCR Integration
Added OCR with pyMuPDF so users can upload photos of textbook pages or handwritten notes. OCR quality is hit or miss depending on the image, so I added text preprocessing to clean up garbage before sending it to the LLM. Handwritten notes are especially rough.
Django Backend
Built the web interface in Django to handle uploads, run the agent, and store quizzes. Used async views so the long-running agent workflow doesn't block everything else. Without that, the whole app would freeze while generating quizzes.
What I Learned
Agent frameworks like LangGraph make complex workflows manageable - the state management handles most of the annoying parts. Iterative prompting works way better than single-shot generation, but you need good evaluation metrics or it just spins in circles. OCR is messier than it looks - most of the work is cleaning up the extracted text. Django's async views are essential for anything that takes more than a few seconds. Designing feedback loops for LLMs is tricky because you need to balance quality improvement with not wasting API calls.