CodeGen - Autonomous Competitive Programming Solver

An end-to-end AI solver built for the Meta Hacker Cup 2025 AI Track.

Links: my-codegen-api2 (production API) and CodeGen-Hacker-Cup-AI-devkit (devkit)

The system behind my Global Rank 10 finish in the Meta Hacker Cup 2025 (AI Track). The AI Track is a parallel competition where participants build AI systems to solve the same algorithmic problems that top human programmers tackle. This project is a purpose-built, end-to-end autonomous solver that accepts a full problem package and returns a validated solution in a single API call.

How it works

  • A FastAPI server (deployed on Render) accepts a problem package (statement, sample input/output and optional images) and generates an optimal Python solution using Google’s Gemini LLM.
  • A complete multimodal pipeline fetches, decodes and passes problem diagrams to the model, since Hacker Cup statements frequently embed graphs and figures served from CDN URLs.
  • The solver enforces the strict competitive-programming I/O format (for example Case #1: 42) by running each candidate against the sample data before returning it.
  • On failure it regenerates iteratively, feeding the error back to the model for up to four attempts, and dedicated endpoints re-run solutions on real competition inputs with LLM-based debugging.

Results

The system solved problems in the upper-medium to hard range, including “Designing Paths” (Round 2) and “Treehouse Telegram” (Round 3), estimated at roughly 2200 to 2500 rating, involving graph algorithms, number theory and careful complexity analysis.

Special mention: the devkit

The CodeGen-Hacker-Cup-AI-devkit documents the full evolution of the solver, from the initial prototypes through the final multimodal API, and ships the reusable client-side automation tools (the CP_GEN toolkit) used to call the deployed API and test solutions during the contest.