Percival Is the Research Data Copilot That Scientists Have Been Begging For

The Macro: Researchers Are Drowning in Data They Cannot Use

Here is a number that should bother everyone in academia: researchers spend roughly 80% of their data analysis time on cleaning, formatting, and wrangling data, and about 20% on the actual analysis that produces insights. That ratio has not changed meaningfully in a decade despite billions of dollars invested in data tools.

The software options for research data analysis have barely evolved. R and Python with Jupyter notebooks remain the default for quantitative research. SPSS is still used in social sciences. Stata dominates economics. Excel, somehow, remains the most common tool for initial data exploration across all fields. Each of these tools requires significant programming knowledge, manual effort for data cleaning, and a level of statistical expertise that many researchers frankly do not have.

The commercial data analysis market has grown enormously, but almost entirely on the enterprise side. Databricks, Snowflake, and dbt are building for data engineering teams at companies. Tableau and Looker are building for business intelligence. Nobody with real traction is building specifically for researchers who need to go from messy CSV files to publishable results.

There have been attempts. Observable tried to modernize notebooks. Deepnote added collaboration to Jupyter. Hex combined notebooks with dashboards. But all of these products assume you already know how to code, what analysis to run, and how to interpret the output. They made the workflow slightly better without solving the fundamental problem: most researchers are not software engineers, and the tools require you to be one.

The AI opportunity here is obvious. If an LLM can write code, it can write the pandas scripts that clean your data. If it can reason about statistics, it can suggest which analysis is appropriate for your research question. If it can interpret results, it can help you understand what your data is actually saying. The question is whether anyone can package that into a product that researchers trust enough to use for work that gets published and peer-reviewed.

The Micro: Two Founders, Free Download, Research Focus

Percival was founded by Kevin Bi and David Zhu, based in San Francisco. They came through Y Combinator’s Spring 2025 batch. The team is three people.

The product, branded as Percy, positions itself as an AI copilot that works inside your existing workspace. You bring your data. Percy writes code to clean and explore it. It suggests analysis tailored to your research objectives. It interprets the results. It can transform and visualize datasets. The idea is that you describe what you are trying to learn from your data, and Percy handles the technical implementation.

That is a big promise, but the execution approach is sensible. Rather than building a completely new environment that researchers have to learn, Percival integrates with existing workflows. There is a free download available, and the product runs through a web app at percy.percivaltech.com with a login. The FAQ and contact pages suggest the team is actively iterating based on user feedback.

The competitive landscape is interesting because the direct competitors are not other startups. They are Jupyter notebooks plus ChatGPT, which is what most tech-savvy researchers already use. Copy your error message into ChatGPT, get a fix, paste it back. That workflow is clunky and context-free, but it works well enough that many researchers have adopted it organically. Percival needs to be meaningfully better than that cobbled-together approach.

Julius AI and Quadratic are probably the closest startup comparisons. Julius lets you upload data and ask questions in natural language. Quadratic is building AI-powered spreadsheets for data teams. Neither is focused specifically on academic research, which gives Percival a positioning advantage if they commit to that niche.

The Verdict

I think Percival is solving the right problem for an underserved audience. Academic researchers are a huge market that commercial data tools have mostly ignored, and the current workflow of Jupyter plus Stack Overflow plus ChatGPT is genuinely bad. If Percy can reliably go from “here is my messy dataset” to “here is a clean analysis with visualizations,” that is a product researchers will pay for and recommend to their labs.

The risk is trust. Academic research has high stakes. A bad analysis does not just look unprofessional. It can lead to retracted papers, damaged careers, and in fields like medicine, real harm. Researchers need to trust that the AI is not hallucinating statistical results or suggesting inappropriate methods. That trust takes time to build and one mistake to destroy.

Thirty days, I want to see what disciplines are adopting Percy first. Social sciences and economics would be strong signals because those fields have the widest gap between data complexity and researcher programming skills. Sixty days, whether the free download is converting to repeat usage or whether people try it once and go back to Jupyter. Ninety days, the strategic question: does Percival go after university-wide site licenses, or does it grow bottoms-up through individual researchers? The former is slower but higher revenue. The latter is faster but harder to monetize. The product concept is strong. The market is ready. The execution needs to earn trust one lab at a time.