An LLM Pair Programmer for Data Analysis, Peer-Reviewed

What it does

LAMBDA is an open-source interface for code-free data analysis driven by natural language. You describe what you want analyzed; a programmer agent generates the Python code, an inspector agent checks and debugs it, and the system loops until the output passes or hits the attempt limit. Everything is exportable to Jupyter Notebook for reproducibility, and the project includes a live web app plus downloadable macOS and Windows clients.

The interesting bit

The project comes out of a statistics department and has actually been published in the Journal of the American Statistical Association, which is more academic credibility than most agent frameworks bother with. The dual-agent setup explicitly separates code generation from verification, and the README notes the team plans to swap the current Gradio UI for OpenWebUI and add Docker support.

Key highlights

• Two-agent architecture: a programmer writes code and an inspector debugs it iteratively.
• Supports OpenAI, Ollama, LiteLLM, and LLaMA-Factory through an OpenAI-style API interface.
• Exports full sessions to Jupyter Notebook for reproduction and further tweaking.
• Includes optional knowledge retrieval for domain-specific analysis tasks.
• Published research with an accompanying survey paper on LLM-based data science agents.

Caveats

• The desktop apps currently need manual kernel setup to run locally.
• Knowledge retrieval is off by default and the module is slated for a ChromaDB refactor.
• The web demo is capped at 200 beta users and requires an invitation code.

Verdict

Data scientists and researchers who want a conversational, iterative coding assistant with academic backing should take a look. If you need a production-grade, fully automated pipeline without human intervention in the loop, this is still experimental.