For research labs
I work with research labs that have more questions than time. I work beside you on your own data, using AI agents to help, until pursuing the answers to those questions is something you can do yourselves.
Scope a pilotScrollThe question is easy to ask, but answering it requires weeks of code for an analysis we'd run once, so we don't do it.
Data-intensive science asks researchers for competence in software on top of their scientific training and areas of specialization. Many labs collect far richer data than they have time to examine or analyze. Implementing even one new analysis method from the literature requires time investment with no guarantee that the findings will be meaningful. It is difficult to justify writing software for a question that will only be asked once. The barrier is time and skills, not curiosity.
AI agents change the calculus of when an analysis is worth doing. The researcher chooses the question, pushes back when the approach drifts, and stays accountable for the findings and the scientific framing. The agents do the building. The fastest way to get started is on your data with someone beside you who has done it before.
New York Scientific Data Summit 2026
Presented as first author at NYSDS, hosted by Brookhaven National Laboratory. Scientists prompt agents to build the views their analysis needs, as they need them. With Dr. Gabriel Salzman and Dr. Everett Meyer.
Read the case studyCenter for the Helping Professions
Six months with Dr. Elizabeth Riley and the research team at a nonprofit working in partnership with the University of Kentucky. The team now extends the workflow on its own.
Read the case studyThere are a dozen published approaches for this, and we only use one.
Agents can search the literature, then implement and tune several approaches so you can see what each one finds. Taken together, they can surface clearer hypotheses than any single method.
Each new analysis could take days or weeks, with no guarantee the results will be useful or scientifically interesting.
When an analysis takes hours instead of weeks, you can try the ones that might not pan out, and keep your attention on the science.
I'd need a custom view of this data to answer that question, and I can't justify the time it would take to write software for it.
Agents make analysis tools fast to build. The more freely you can explore, the more likely you are to find interesting science, and the better you understand your experiment and data along the way.
The answers are in our records somewhere, but nobody can read them all.
Language models can classify and extract at scale, checked against your experts' judgment, inside the environment your data has to stay in.
The researcher chooses the question, and we discuss how we'll know whether the results are meaningful. Agreeing on a way to verify them up front lets us use agents to explore without relying on them alone for correctness.
Agents search the literature, implement and tune approaches, and build the visuals and analysis the question needs. I help direct them, and we push back whenever the approach drifts from the intent.
Every candidate result is verified by the scientist before it counts. Building that check into the analysis is what makes agent-written code trustworthy enough to be useful.
You prompt the agent to explore and run the next analysis yourself.
Start small, see results on your data, then decide how far to take it.
We investigate a question using data you already have. I support you closely at the start, and by the end you're increasingly independent, directing an agent to carry out the analysis.
An analysis approach built around your data and questions, and coaching for each researcher until they can run, verify, and extend it themselves.
A standing working relationship to take the work further: reviewing approaches, working through hard problems together, and weighing in high-leverage decisions.
The researcher chooses the question, pushes back when the approach doesn't match their intent, and decides what to do next. They set the direction and ground the analysis in the science.
Each analysis method provides a different look into the data. You can compare several and follow what they agree on when agents make implementing a method cheap.
Agent-written analysis is only useful if you can verify it. We agree on the checks before the work starts, and the scientist verifies every result.
Success is your lab running the next analysis on its own, with code and methods it owns and knows how to extend.
I have found tremendous educational value in working with Dan on several projects across disciplines. He models a unique ability to elicit, frame, and target the core features of any problem. Dan taught me to build such a skillset and use it, with the help of AI tools, to find elegant and rigorously validated solutions using powerful computational approaches outside my domain of expertise.
Dan came into the project without deep content expertise, and he worked hard to understand the full scope and nature of the problem we were trying to solve. His strength is finding responsive, creative technical solutions using software development, machine learning, and AI, and he helped us create a workflow to conduct analysis at scale in a highly secure environment, something that simply was not possible before in our work. He worked with our non-technical team collaboratively, with a great deal of patience and curiosity, and helped us understand our own research questions in a new way.
Tell me about your data and the question you've been waiting to answer. We'll figure out on a short call whether a pilot is the right place to start.