MathModelAgent: an AI team that writes your modeling paper in one hour

What it does

MathModelAgent orchestrates specialized LLM agents (modeler, coder, writer) to tackle mathematical modeling problems end-to-end. It analyzes the problem, builds models, writes and executes code via Jupyter or cloud interpreters like E2B and Daytona, then outputs a formatted paper ready for submission. The whole thing is wired together as a custom workflow without heavy agent frameworks, keeping API costs down.

The interesting bit

The project treats a math modeling competition like a small software sprint with human-in-the-loop checkpoints. The author claims the goal is compressing three days of work into one hour—though the README candidly admits current output won’t actually win national competitions yet. The four-layer fault tolerance (retry → fallback → evaluator shadow → feedback rerun) is ambitious on paper, even if much of it remains TODO-stubbed in the codebase.

Key highlights

• Multi-LLM support via LiteLLM; each agent can use a different model
• Local Jupyter or cloud code interpreters (E2B, Daytona) with notebook persistence
• Web search via OpenAlex for real data gathering (Tavily API planned but not implemented)
• RAG knowledge base scaffolding with ChromaDB + rerank for retrieving modeling methods and templates
• HIL (human-in-the-loop) with six decision actions at key workflow nodes
• Docker and local deployment options; WebUI and CLI both available

Caveats

• Several marquee features are partially or entirely unimplemented: HIL workflow integration, RAG core retrieval logic, fallback hand-off, evaluator feedback loop, and Tavily web search all have TODO comments or placeholder configs
• The author explicitly warns the project is experimental, buggy, and updated sporadically
• English support and LaTeX template integration are still pending, limiting use for international competitions like the MCM/ICM

Verdict

Worth a look for Chinese math modeling competitors who want to experiment with AI-assisted workflows and don’t mind rough edges. Skip it if you need production reliability, English output, or a fully realized agent framework—the README is admirably honest about what’s still duct tape and dreams.