Cooked

The idea behind this project is to make learning competitive programming less difficult and frustrating, especially for those who practice alone.
This application aims to support learners by integrating an LLM agent that provides hints, explains coding errors, and guides users step-by-step toward solving problems.

To achieve this, we first deployed Ollama on Google Cloud Run, which was our first experience using this platform.
This setup allowed us to create an API capable of calling an LLM.
We chose Gemma 3 (4B) for its balance between model size and efficiency.
The instance ran on an NVIDIA L4 GPU, with 16 GB of RAM and 8 CPU cores.

Next, we developed a Flask web application with a clean and simple interface.
The home page displays the problem sets stored in our database, since scraping problems from Codeforces (https://codeforces.com/)
is both challenging and restricted by the website.
Additionally, the Codeforces API does not provide full problem details such as statements, inputs, outputs, and constraints.
Therefore, we manually curated a database of problems.

Because the copied problems were not well formatted, the LLM reformats them into Markdown before displaying them to the user.
It also provides three initial hints, each offering progressively more guidance.
Users can request additional hints at any time.

We also implemented a working VS Code instance on the right panel of the interface, which allows users to run their code on the Google Cloud Run server.
The application then verifies the output and compares it to the expected result.
Finally, users can directly interact with the LLM by sending custom messages for clarification or help.