Uncookd
What Inspired Us
We were inspired by a fundamental problem with modern AI tutors: they're oracles, not companions. They give you the answer, which encourages shallow learning. We wanted to build a responsible AI mentor that fosters critical thinking - a virtual study group where students could reason together with AI, not just get answers from it.
How We Built It
Our project is a real-time, multi-agent virtual classroom built on a modern tech stack:
- Frontend: Next.js, React, integrated with tldraw for a live, interactive whiteboard.
- Backend: Python, FastAPI, CrewAI
AI Core
CrewAI to orchestrate a team of specialised AI agents:
- Socratic Mentor: Guides the user with questions, never giving the answer outright.
- Domain Analyst: A subject-matter expert that reasons with the user and plots graphs directly onto the whiteboard.
- Critical Thinker: A "devil's advocate" agent that challenges assumptions to promote deeper thinking.
What We Learned
Agents > A Single Model:
A multi-agent system is far more robust and responsible. By separating concerns (e.g., Socratic guidance vs. critical analysis), we built a system that is more auditable, safer, and avoids the "do-it-all" pitfalls of a single LLM. The pivotal moment was realising LLMs are best used as reasoning engines that delegate actions to other agents and tools.
