-
Initial Look
-
Raw
-
Sizzling
-
Cooking
Inspiration
Over the last year we noticed something strange happening around us. People were no longer asking AI how to learn; they were asking it to do the thinking for them. Students were submitting essays they never understood. Friends were asking AI what to text someone. People were even using AI to decide how they felt about situations. AI had become a replacement for cognition rather than a tool for it. The moment that really motivated this project was simple: we realized modern AI is optimized for correct answers, not for human growth.
We asked a different question: What if an AI refused to give you the answer… until you showed effort?
COOKED was born from that idea - not an AI assistant, but an AI cognitive trainer.
What it does
COOKED is a Socratic AI tutor. Instead of solving your problem, it guides you to solve it yourself.
When a user asks a question, COOKED does not immediately respond with an answer. Instead, it: Detects the task type (math, writing, or conceptual explanation) Asks adaptive leading questions Measures learning effort Unlocks help progressively
We built an Effort Meter that scores how much thinking a user actually does.
The system evaluates: -time spent before responding -number of revisions (backspaces) -attempt count -reasoning structure -use of examples or explanations
How we built it
React, Node.js, Vite- UI
Python, FastAPI - Backend
Gemini - LLM used
Challenges we ran into
1) The LLM Gemini has a very small free-trial quota, so there is a rate limit for the program.
2) The Over-helpful AI Problem: LLMs are trained to be helpful. Our entire project required the AI to sometimes refuse to help. We had to design prompts that encouraged learning but did not frustrate users.
3) Effort Detection: Users quickly tried to game the system by: spamming long messages, pasting random text, repeatedly requesting final answers. We solved this using saturating scoring functions and structure detection rather than simple length counting.
4) Human Factors: A major realization was that some users genuinely need guidance, not restriction. We are looking to add a supportive interaction style so the system encourages thinking without abandoning the learner.
What's next for Cooked
We want to expand COOKED into: -study-habit coaching -learning analytics dashboards -adaptive difficulty tutoring -mental-wellbeing supportive mode -integration into LMS platforms Our goal is an AI that acts less like a search engine and more like a teacher.
Log in or sign up for Devpost to join the conversation.