Data Science Agent

The AI data scientist that shows its work.

Ask questions about CSV files and databases in natural language. DSA runs SQL, statistics, machine learning, and visualization, then returns reproducible artifacts with claim-level evidence.

Quickstart · See it in action · Case studies · Evaluation · Docs · Roadmap

Why DSA?

Most AI data-analysis tools stop at the answer.

Data Science Agent is built to preserve the path from a claim back to the computation and dataset that produced it.

Question
   ↓
Executed analysis
   ↓
Evidence
   ↓
Claim
   ↓
Reproducible report

That means an analysis can be inspected, audited, reproduced, and challenged instead of accepted on confidence alone.

If an AI makes a data-science claim, you should be able to inspect how it got there.

⚡ 60-second quickstart

Install:

pip install jack-data-science-agent

Run the built-in demo:

dsa demo

Analyze your own data:

dsa analyze sales.csv \
  --task "Which factors explain revenue, and are the effects statistically significant?"

Python 3.12+ is required.

A successful run produces more than chat output:

analysis run
├── report.md
├── experiment.json
├── evidence_graph.json
├── analysis.ipynb
└── reproduce.sh

See it in action

_{Question → profile → plan → tools → evidence → critic → reproducible report.}

Start with 3 flagship workflows

Browse all 8 verified case studies →

These are verified repository workflows intended to show product behavior and evidence artifacts. They are not presented as independent real-LLM leaderboard results.

What makes DSA different?

1. Claim-level evidence

Supported findings can be connected to the analysis that produced them:

Claim
└── statistical result
    └── executed tool
        └── parameters
            └── dataset SHA-256

2. Reproducibility by default

DSA records run metadata and reproduction artifacts instead of treating the final prose answer as the only output.

3. Evidence-aware critique

A critic stage checks whether conclusions go beyond available evidence. For example, causal language can be rejected when the underlying analysis only supports correlation.

4. Real data-science tools

Core workflows can coordinate:

• dataset profiling and exploratory analysis
• SQL with DuckDB
• statistical testing
• regression and classification
• forecasting and feature importance
• visualization
• report generation

5. One runtime, multiple interfaces

Python SDK

import asyncio
from data_science_agent import Agent

async def main():
    result = await Agent().analyze(
        "sales.csv",
        "Which region drives the most revenue, and is the trend statistically significant?",
    )

    print(result.report_markdown)

asyncio.run(main())

Inspect evidence programmatically:

for evidence in result.evidence:
    print(evidence.claim)
    print(evidence.source_id)
    print(evidence.result)

Case study gallery

The repository includes verified end-to-end examples across business analytics, churn, forecasting, marketing, finance, public statistics, data quality, and classification.

Browse the Case Study Gallery →

A useful way to evaluate DSA is to start with the question you would ask about your own dataset, then inspect the generated report and evidence trail.

Evaluation

DSA is developed against versioned benchmark suites rather than relying only on hand-picked demos.

The repository currently contains deterministic benchmark and evaluator-validation results, including frozen task suites and reproducibility checks. These results are useful for regression testing and validating the evaluation harness, but they should not be interpreted as an independent comparison of real LLM model quality unless the run identifies a real model/provider and reproducible configuration.

The current public result registry includes a stub/small validation run. It is intentionally labeled as such so test-harness scores are not confused with real-model performance.

• Evaluation methodology
• Benchmark documentation
• Reproducible result registry
• Research & limitations

Next evaluation gate: run the merged credentialed four-way real-model smoke workflow, review all raw artifacts for reproducibility, and only then prepare any full-catalog comparative result for publication.

Architecture

flowchart LR
    A["Natural-language question"] --> B["Planner"]
    B --> C["Data-science tools"]
    C --> D["SQL / Statistics / ML / Visualization"]
    D --> E["Evidence graph"]
    E --> F["Critic"]
    F --> G["Evidence-backed report"]
    G --> H["Reproduction artifacts"]

The runtime uses a LangGraph-based orchestration layer over typed data-science tools, with local-first data operations built around Python, DuckDB, Polars, SciPy, scikit-learn, and Matplotlib.

DSA vs. typical AI data analysis

DSA is not intended to be only another natural-language interface to a dataframe.

Its focus is verifiable, reproducible AI data science.

Documentation

Start here:

• Getting Started
• SDK & API Reference
• Evaluation
• Benchmarks
• MCP
• Research & Limitations
• Case Studies
• Roadmap
• Changelog
• Releases

Development

git clone https://github.com/Jackxiaozhiren/data-science-agent.git
cd data-science-agent
uv sync --dev
uv run dsa demo
uv run pytest

Static checks:

uv run ruff check .
uv run mypy .

Contributing

Contributions are welcome, especially:

• real-model benchmark baselines
• reproducibility failures
• new datasets and benchmark tasks
• statistical validation improvements
• data-science tools and plugins
• case studies and documentation

New contributor? Start with the good first issue queue.

See CONTRIBUTING.md, ROADMAP.md, and CODE_OF_CONDUCT.md.

Citation, security, and license

• Academic use: CITATION.cff
• Security reports: SECURITY.md
• License: MIT