K-Dense-AI/karpathy

What it solves

It provides an agentic Machine Learning Engineer capable of training state-of-the-art ML models. It simplifies the process of setting up a specialized environment where an AI agent can leverage scientific workflows and ML libraries to perform engineering tasks.

How it works

The project uses the Claude Agent SDK and Google ADK to implement an agent that operates within a dedicated sandbox directory. This sandbox is populated with a collection of over 130 ready-to-use "Scientific Agent Skills" and a Python virtual environment containing essential ML packages like PyTorch, transformers, and scikit-learn. Users interact with the agent via a web interface provided by the ADK.

Who it’s for

It is designed for developers and researchers who want to use AI agents to automate the training of machine learning models and implement scientific research workflows.

Highlights

• Integrates with the Claude Agent SDK and Google ADK.
• Leverages a library of 130+ specialized Scientific Agent Skills.
• Automatically configures a sandbox environment with pre-installed ML libraries (PyTorch, scikit-learn, etc.).
• Supports integration with tools like Cursor, Claude Code, and Codex.