plexe: Build Machine Learning Models from Prompts

Overview

Plexe is a Python tool that uses AI agents to build machine learning models from a natural-language intent and a tabular dataset. It automates task analysis, model search, evaluation, and packaging, addressing the setup and iteration work involved in turning a dataset into a deployable model.

It is suited to developers and data teams who want an automated starting point for tabular ML, rather than a no-code interface for every stage of the data science process. Plexe requires access to LLM providers and produces a self-contained model package; a managed cloud service is also available.

Key Features

• Uses 14 specialized agents across a six-phase workflow to analyze data, select evaluation metrics, search for models, and assess results.
• Builds tabular models with XGBoost, CatBoost, LightGBM, Keras, and PyTorch, subject to installed dependencies.
• Accepts datasets in Parquet, CSV, ORC, or Avro formats and a natural-language modeling intent.
• Packages trained artifacts, inference code, schemas, configuration, and evaluation reports for deployment outside Plexe.
• Supports LLM routing through LiteLLM, with OpenAI and Anthropic models actively tested according to the README.
• Offers Docker images, YAML configuration, and a Streamlit dashboard for experiment results.
• Provides an integration interface for custom storage, tracking, and deployment infrastructure.

Use Cases

• A data scientist wants to quickly establish a model baseline for a tabular prediction task and explore candidate approaches.
• A machine learning engineer wants to automate repetitive model-search and evaluation steps before refining the resulting model.
• A small team with limited ML engineering capacity wants to turn a structured dataset and a clear prediction goal into a packaged model.
• A platform team wants to experiment with agent-driven model building while routing different workflow steps to different LLM providers.

Getting Started

Install Plexe and set the provider credentials needed by your configuration:

pip install plexe
export OPENAI_API_KEY=<your-key>
export ANTHROPIC_API_KEY=<your-key>

See the README for dataset requirements, CLI options, optional dependencies, and configuration details.

Alternatives

• AREX-Skill: AREX-Skill equips coding agents with reusable ML skills, rather than providing an end-to-end workflow for training and packaging tabular models.

Considerations

• The tool requires API credentials for the LLM providers used by its agents. The README says other LiteLLM providers may work, but only OpenAI and Anthropic models are actively tested.
• Python support is limited to versions >= 3.10 and < 3.13.
• The workflow targets tabular datasets. Additional model frameworks and platform integrations require optional dependencies.
• Automated model selection does not remove the need to validate suitability, data quality, and performance for the intended application. The README includes optional held-out evaluation.
• The README describes a managed cloud service as an alternative to running the Python package; deployment and usage details are on the project site.