Debug Demolition

Inspiration

With the growing power of large language models that are increasingly proficient at coding, it is more important than ever to distinguish hallucinations from accuracy. Thus, proper and efficient debugging has become an invaluable skill. Programming competitions/platforms such as Leetcode/ICPC are fun, but there are few alternatives in other areas.

What it does

Matches users into rooms of 2+ to submit a corrected version of a flawed programming solution generated by AI. They are judged based on accuracy and speed under a given time limit. Afterwards, users' statistics and history, including ELO rating, are updated.

How we built it

We use Websockets to connect and pair players into a room, which syncs data and actions with all connected sockets. Deepseek was used to generate incorrect problem statements from known correct CSV data, and Python/Javascript was used to parse input and output.

Challenges we ran into

Using web sockets were consistently quite tricky. Given the loose typing of Javascript and the difficulty of perfectly inspecting socket state, we frequently needed to use print statements for debugging. It also took us a while to figure out structural problems, such as how to persist sockets that rendered under the router for navigation, while ensuring that any changes wouldn't close the current instance (solution: use a singleton).

Accomplishments that we're proud of

Being able to bridge the network connections with varying API calls with levels of parsing.

What we learned

Typed languages are much nicer for catching errors. LLMs often omit small sections of code over large queries/tokens. More often than not unusual errors can be traced to unexpected causes (ex: hitting the token limit -> no response on await).