jihe520/MathModelAgent

What it solves

MathModelAgent automates the entire process of mathematical modeling for competitions. It transforms a multi-day manual effort into a streamlined workflow that handles problem analysis, mathematical modeling, code implementation, and the final drafting of a professional, submission-ready paper.

How it works

The project uses a multi-agent system consisting of specialized roles (such as a Modeler, Coder, and Writer) that can be powered by different LLMs via LiteLLM. It operates as a set of "SKILLS" that can be driven by harnesses like Claude Code or Codex. The workflow includes:

• Code Execution: A built-in Code Interpreter (local Jupyter or cloud-based via E2B/Daytona) to run simulations and generate data.
• Document Generation: Uses Typst templates to produce high-quality PDF papers in formats compatible with major competitions (e.g., MCM/ICM).
• Knowledge Integration: A RAG-based knowledge base containing modeling standards and decision trees to reduce hallucinations.
• Verification: A 9-step automated acceptance process that checks for text leaks, numerical consistency, and compilation errors.
• Human-in-the-Loop: Allows users to approve, edit, or regenerate results at critical nodes.

Who it’s for

Students and researchers participating in mathematical modeling competitions who want to automate the tedious parts of problem-solving and paper formatting.

Highlights

• End-to-End Automation: A single command can trigger the entire pipeline from analysis to PDF generation.
• Professional Templates: Includes 17 Typst templates for various national and international competitions.
• Flexible Deployment: Available as a standalone desktop app, Docker container, or local installation.
• Extensible Architecture: Skills can be called individually or combined into custom workflows.