MLE-Agent: Your intelligent companion for seamless AI engineering and research.

💌 Fathers' love for Kaia 💌

Overview

MLE-Agent is designed as a pairing LLM agent for machine learning engineers and researchers. It is featured by:

• 🤖 Autonomous Baseline Creation: Automatically builds ML/AI baselines.
• 🔍 Arxiv and Papers with Code Integration: Access best practices and state-of-the-art methods.
• 🐛 Smart Debugging: Ensures high-quality code through automatic debugger-coder interactions.
• 📂 File System Integration: Organizes your project structure efficiently.
• 🧰 Comprehensive Tools Integration: Includes AI/ML functions and MLOps tools for a seamless workflow.
• ☕ Interactive CLI Chat: Enhances your projects with an easy-to-use chat interface.
• 📊 Weekly Report: Automatically generates detailed summaries of your weekly works.

mle_v030.mp4

Milestones

• 🚀 09/10/2024: Release the 0.4.0 with new CLIs like MLE report, MLE kaggle, MLE integration and many new models like Mistral.
• 🚀 07/25/2024: Release the 0.3.0 with huge refactoring, many integrations, etc (v0.3.0)
• 🚀 07/11/2024: Release the 0.2.0 with multiple agents interaction (v0.2.0)
• 👨‍🍼 07/03/2024: Kaia is born
• 🚀 06/01/2024: Release the first rule-based version of MLE agent (v0.1.0)

Get started

Installation

pip install mle-agent -U
# or from source
git clone git@github.com:MLSysOps/MLE-agent.git
pip install -e .

Usage

mle new <project name>

And a project directory will be created under the current path, you need to start the project under the project directory.

cd <project name>
mle start

You can also start an interactive chat in the terminal under the project directory:

mle chat

Use cases

Generate Work Report

MLE agent can help you summarize your weekly report, including development progress, communication notes, and to-do lists.

cd <project name>
mle report

Then, you can visit http://localhost:3000/ to generate your report locally. Alternatively, you can directly try our deployed service at https://workspace.repx.app/ to generate reports with more third-party extensions (e.g., Zoom, Google Calendar) supported.

Start with Kaggle Competition

MLE agent can participate in Kaggle competitions and finish coding and debugging from data preparation to model training independently.

cd <project name>
mle kaggle

Roadmap

The following is a list of the tasks we plan to do, welcome to propose something new!

🔨 General Features

• Understand users' requirements to create an end-to-end AI project
• Suggest the SOTA data science solutions by using the web search
• Plan the ML engineering tasks with human interaction
• Execute the code on the local machine/cloud, debug and fix the errors
• Leverage the built-in functions to complete ML engineering tasks
• Interactive chat: A human-in-the-loop mode to help improve the existing ML projects
• Kaggle mode: to finish a Kaggle task without humans
• Summary and reflect the whole ML/AI pipeline
• Integration with Cloud data and testing and debugging platforms
• Local RAG support to make personal ML/AI coding assistant
• Function zoo: generate AI/ML functions and save them for future usage

⭐ More LLMs and Serving Tools

• Ollama LLama3
• OpenAI GPTs
• Anthropic Claude 3.5 Sonnet

💖 Better user experience

• CLI Application
• Web UI
• Discord

🧩 Functions and Integrations

• Local file system
• Local code exectutor
• Arxiv.org search
• Papers with Code search
• General keyword search
• Hugging Face
• SkyPilot cloud deployment
• Snowflake data
• AWS S3 data
• Databricks data catalog
• Wandb experiment monitoring
• MLflow management
• DBT data transform

Contributing

We welcome contributions from the community. We are looking for contributors to help us with the following tasks:

• Benchmark and Evaluate the agent
• Add more features to the agent
• Improve the documentation
• Write tests

Please check the CONTRIBUTING.md file if you want to contribute.

Support and Community

• Discord community. If you have any questions, please ask in the Discord community.

Star History

License

Check MIT License file for more information.