I'm bad at LeetCode.
@kamathematic5.9K followers
151,154 views
55 Comments
Comments from the original post on X (Twitter)

@kamathematic
Throw me a star if you can :) Built using @bunjavascript, @nextjs, @aisdk, @honojs, and @Cloudflare sandboxes + workflows
45 likes

@kamathematic
@nextjs @aisdk @honojs @Cloudflare Damn, in 24 hours! This was just a side project I built for myself over Thanksgiving and thought I’d share. Yes, it’s buggy, fixing 🫡🛠️
8 likes

@will__ye
diabolical name
15 likes

@kamathematic
if they ever find out I’m cooked
4 likes

@hackerrank
@will__ye We found out. Great concept though, will try it out.
1 like

@kamathematic
@will__ye lmao i meant if the clankers ever find out 😭

@sp2hari
Interesting name. Agree with you on building a platform for learning the concepts instead of memorizing solutions. Happy to connect sometime to understand more. Drop a note to hari @ hackerrank if you are interested.
5 likes

@kamathematic
Honored to have made it to your feed! Emailed you :)
1 like

@aakashreddy
This works till you see LLM generate some bullshit test cases and logically ill sounded bullshit questions. It can easily fuckup even if you ask a leetcode question and ask it to diverge from editorial solution, and it goes absolutely bonkers
2 likes

@kamathematic
absolutely! far from perfect, but there’s a few guardrails you can put in place. for example, LLMs can def hallucinate entire test cases, but you can use LLMs to describe test cases (“a large nondecreasing array of prime numbers”), then use that description to generate code to generate test case inputs. run the solution through those inputs to get expected outputs. the tricky part is in a hallucinated solution, but there’s likely ways to use LLM as a judge here. this is just standard leetcode, not IOI, so shouldn’t be too crazy
1 like

@aakashreddy
Guard rails take you to a particular level, but let us say for a 2D array of arrays, it just gives bullshit output for a given input, and you ask it to check again it gives a different answer, Only way to solve this is generate test cases using the correct code, which is not confirmed to be correct if written by a LLM. But could definitely be an accelerator for this purpose

@kamathematic
yeah so the LLM generates code to generate appropriate inputs and then a solution. it runs the inputs through the solution to get the expected output, preventing the case in which you get different expected outputs every time. the only real problems are in bad input generation code (not likely, pretty trivial to do write code to generate a string/array/number that fits a certain constraint) or a bad solution. bad solution is decently mitigated with LLM as a judge. all in all practical worst case is you can just regenerate a problem haha

@aakashreddy
Yep, only case this remotely works or illuses to work, is when we run the solution for generated test cases in our execution environment, rather LLM assuming the output

@sashimikun_void
I am also bad at Leetcode , we should chat! haha
2 likes

@vatsandev
This is cool! Would love to check this out, maybe a quick idea would be to have a section which lets this AI help you understand Advent of Code?
1 like

@kamathematic
Ooo super interesting! Can look, but also wouldn't mind if you opened a PR!
1 like

@vatsandev
Yeah something like a plug-in where in you can choose from previous year questions and on going questions. Will take a look asw, gave it a star!
1 like

@AnkitaDuri
@MehtaDontStop 👏
1 like

@whoiskatrin
that’s cool! 👌🏼
1 like

@kamathematic
thanks! first time using cloudflare, you guys are killing it!
2 likes
























