So I have been working on a project and I think it's finally ready.
CodeCoach AI is a recommendation system that provides you the topics one might need to practice in order to break away from their current ratings.
https://code-coach-ai-steel.vercel.app/
The working is explained in the ReadMe on github which is accessible from the bottom of the home page. But the basic premise of the project is that instead of going through the route of checking the user submissions to identify the weak topics, I thought that if a person is stagnant on a rating, he ought to be doing what those about 100 to 200 rating points above him have done, since they have broken the barrier the user is stagnant on.
So I first made a pre-compiled collection of users and their submissions, stored in form of rating bands. from 800-900 to 2400-inf. Each rating band has 300 random users of that rating, and for each user, we take the last 500 submissions. Then , lets say if a user is of rating X, we compare his submissions with the compiled data of users in band x+100 to x+200. This tells us how our user fares against his peers. This gives a collection of various topics (sortings, binary search, greedy, dp, etc) , aranged by how good the user performs in those topics compared to peers. This is sent to an LLM whose only job is to present our json object in natural language. All the calculations are deterministic, not done by LLM.
Then for the top 3 worst topics,we return the top 5 unsolved problem of those topics in range which is attempted most commonly by the members of the rating band. This data us also presented in form of graphs.
I would love users here to test this and help me with any suggestion they have for improving the project. That would benefit me a lot. Thank you
It actually seems pretty accurate for me, I tried it out.
Edit: I will add that the topic competency is inflated if a user has solved more problems, which is to be expected, but solving more problems will not always fix a plateau.
Thanks for trying man . And yes , I wanted to handle that. So I tried using metrics like attempt ratios , coverage , solve rate to balance the volume. But yeah, it can influence the perceived competency of a user. I would definitely wanna refine it in future iterations
Thinking quickly, siddharth1119sid constructs a homemade problem recommender, using only some rating bands, an LLM, and a problem recommender.
"solve problems above your level" doesn't need to be this complicated vro
Everyone knows "solve harder problems vro". Which topics to focus for the best ROI? What difficulty to choose? That's what was in my mind while making this. Not everyone has the intuition of an expert after all vro
literally don't focus on anything. any problem you solve above your skill level is good. and i mean ANY. the only times you need to focus on a specific topic are:
you're extremely weak in one. like if you're an expert in graph problems but can't find the prime factorization of an integer. (if you don't know if you're in this category, you're not in it)
you're training for a team contest and you specialized in this topic so the rest of the team can focus on other things. and usually you pick the topics you're already good at.
and in both of these cases, you can just filter for problems with that tag.
Good work bro!
Something went wrong