Inspiration

A multi-MCP, multi-agent coder that specializes in winning kaggle competitions.

With years of experience between the team members on AutoML and LLM training, we wanted to build a freer, more powerful machine learning agent leveraging the latest and greatest advancements in foundational language models. Think ClaudeCode for Kaggle.

What it does

Integrates with Claude Desktop to go from file path to trained machine learning model with metrics. CSV format is assumed for now, but other than that there are few assumptions on the dataset, task, etc.

How we built it

Multiple MCP servers with various specialties: data analysis, preprocessing, and writing training code with self healing.

Challenges we ran into

Agent tends to launch off longer form training runs (e.g. 1hr+) which would be correct in a "normal" setting but is too slow of an iteration speed for a < 8 hour hackathon

Accomplishments that we're proud of

First time working with MCPs for each of the team members!

Built With

Share this project:

Updates