Keiji AI LogoKeiji AI

Research · Nature Biomedical Engineering 2026

DSWizard

Research on biomedical data science agents. DSWizard is an AI agent for biomedical data science that begins with an analysis plan and refines it before writing code. The paper first benchmarks how reliable language models are at this work.

Who created DSWizard?

DSWizard was written by Keiji AI co-founders Zifeng Wang, Benjamin Danek and Jimeng Sun, who is Keiji AI's CEO, with Ziwei Yang and Zheng Chen. The paper lists Sun with a Keiji AI affiliation.

What the paper found

  • Benchmarked eight proprietary and eight open-source language models on 293 coding tasks from 39 studies across seven research areas. Overall accuracy was below 40%.
  • The plan-first agent reached 74% accuracy.
  • In a study with five medical researchers, users completed over 80% of the analysis code for three studies.

How DSWizard relates to TrialMind

It is research, not a TrialMind module or a separately sold product. Published methods like this one inform how TrialMind works; TrialMind is the product Keiji AI sells.

See how TrialMind handles this work: Real-world evidence.

The paper

Zifeng Wang, Benjamin Danek, Ziwei Yang, Zheng Chen, Jimeng Sun. Making large language models reliable data science programming copilots for biomedical research. Nature Biomedical Engineering, 2026.

Details checked against the published record on 2026-11-01.

Frequently asked questions

Who created DSWizard?
DSWizard was written by Keiji AI co-founders Zifeng Wang, Benjamin Danek and Jimeng Sun, who is Keiji AI's CEO, with Ziwei Yang and Zheng Chen. The paper lists Sun with a Keiji AI affiliation.
Is DSWizard part of TrialMind?
No. It is research, not a TrialMind module or a separately sold product. Published methods like this one inform how TrialMind works; TrialMind is the product Keiji AI sells.
What did the DSWizard paper find?
Benchmarked eight proprietary and eight open-source language models on 293 coding tasks from 39 studies across seven research areas. Overall accuracy was below 40%. The plan-first agent reached 74% accuracy. In a study with five medical researchers, users completed over 80% of the analysis code for three studies.

All publications

Every paper sits under its primary research area.