GPdeTox

Inspiration

Developers today spend hours monitoring and talking to agents to write code for them. While this is very efficient, we wanted to make something that brings developrs back to their roots by making them write code again, but with a new creative agentic twist. Inside GPdeTox, you can experience what its like to be an AI agent being prompted by a passive aggresive and tired senior engineer. Often we are told to review AI-generated code. But as people forget how to code, it becomes much easier for bugs to slip by code reviews.

What it does

GPdeTox puts you inside a Cursor-like application, supplying you with a terminal to talk to your human overseer, as well as your coding enviromnent. You are given a task to solve that mimics real development challenges like implementing a UI feature or finding a bug. While you work on the task, the AI impatiently waits for you, while keeping track of how long you stay idle for, your approach and much much more. As an agent, it is your job to make sure the human is satisfied with your work. For full immersion enable audio to hear the developer give you feedback courtesy of ElevenLabs.

How we built it

A deterministic Java TriggerEngine watches editor telemetry and decides when the interviewer should speak. Idle too long, a suspicious paste, a submit. Only then does a model call happen. That keeps cost bounded and makes OpenAI's prompt caching pay off: every call shares a stable prefix (the interviewer's voice, the problem, the rubric) with just a small volatile tail changing. Stack: Spring Boot on Java 25, SQLite, the OpenAI Java SDK for structured-output calls, React + Typescript + Vite with Monaco as the editor. One Maven build, one runnable jar. ElevenLabs API

Challenges we ran into

Lack of diversity in AI responses, which was fixed by fine tuning the system prompt over 20 times.

Accomplishments that we're proud of

We succesfully made a snarky and cruel AI code reviewer We gave him a voice by linking him to ElevenLabs We created something that we could see ourselves using

What's next for GPdeTox

Different personalities for code reviewer Allow user to influence what type of question they get Improve the prompt more so it is even more human sounding Implement different game modes