For more than ten years I did the serious version of this work.
A PhD in epidemiology. National Director of Epidemiology for a ministry of health. National focal point to PAHO for surveillance and the International Health Regulations. Project coordinator on a PAHO/WHO programme. Lead consultant on an eight-country study across Central America and the Dominican Republic. Graduate faculty teaching research methodology, and examiner on a dozen master’s theses.
I published. I advised governments. I taught causal inference to people who went on to teach it themselves.
And I kept walking into the same wall.
The wall
Health organizations sit on mountains of data and make decisions with a fraction of it.
Not because the people are careless — I have worked with some of the most careful people I know. Because of a structural gap that nobody’s job description covers.
The methods exist. The papers exist. Someone published the right approach to that surveillance problem in 2019, and it is rigorous, and it is sitting in a PDF behind a paywall in a language the person who needs it does not read comfortably, with an R script in a supplementary appendix that has not run since the package was updated.
What does not exist is the person who turns rigorous science into working software.
Not a data analyst who produces a report. Not a software vendor who produces a dashboard. Someone who can hold both ends: who knows why the adjustment set matters and can also write the deployment pipeline that keeps the calculation correct in production at three in the morning.
The two halves rarely meet
I have sat on both sides of this table.
I have watched technology vendors present health systems with beautiful products built on a methodological error that nobody in the room was equipped to catch. Confidence intervals computed as if the sample were simple random when it was clustered. Risk scores adjusted for a mediator. Models validated on the same data that trained them, described in language that made it sound otherwise.
And I have watched excellent epidemiologists produce analyses that were completely correct and completely unusable — a 200-page report that answers the question the ministry asked eleven months ago, delivered as a static document that cannot be re-run when the next outbreak has different parameters.
Neither side is at fault. They were trained for different things, and the gap between them is not a skills gap that a workshop fixes. It is a gap in who is willing to be accountable for both halves at once.
So I built the thing I could not hire
IAM-Studio is an AI studio born inside public health rather than imported from technology.
The distinction is not marketing. It shows up in what we refuse to ship.
We build causal inference tooling, clinical and public health decision support, surveillance systems and research infrastructure. What we do not build is a generic chatbot with a health sticker on it — the category that has done more to exhaust the goodwill of clinicians than any other in the last three years.
A few principles that have held up:
The formal engine decides, never the language model. In IAM-CausalKit, a language model may propose a graph, but the adjustment set is verified by a deterministic d-separation algorithm. If the algorithm cannot prove it, the tool says so. No causal claim is ever emitted by an LLM.
Nothing that decides a number lives in the interface. In IAM-InvestKit, the calculation engine is fully decoupled from the UI and carries 343 tests that run inside the Docker build. A broken calculation cannot reach production, because the build fails first.
Test the invariants, not the screen. EPINEXUS has a test suite written over scientific invariants — properties that must hold for the result to be epidemiologically meaningful — rather than over the interface that displays them.
Measure cost like latency. In one product we instrumented AI cost per call from day one. It came to $0.00045. That number retired a planned usage cap that would have limited growth to save fractions of a cent — a decision we would have made on vibes otherwise.
The day job is not a side note
I still direct the Centro Coordinador de Salud para el Bienestar in Tamaulipas, Mexico.
People occasionally suggest this is a conflict, or at least a distraction. It is the opposite. It is the quality control.
Every tool I build gets tested against a question I cannot avoid: would this actually survive here? In a clinic with intermittent connectivity, thin staffing, a data capture form somebody filled in at the end of a twelve-hour shift, and a decision that has to be made this week regardless of whether the data is clean?
That is why the measles pathway we built for first-level health workers is a single offline file. Not because offline-first is an elegant architecture — because the connection drops, and the patient is still in front of you.
You cannot get that instinct from a requirements document. You get it from having been the person who has to answer for the outcome.
What I am actually trying to do
I am not trying to convince anyone that AI will transform health systems. That claim has been made loudly enough by people with more funding and less exposure to the consequences.
I am trying to demonstrate something narrower: that it is possible to build software for health that a methodologist would defend and an operator would actually use. That rigour and shipping are not in tension. That the reason most health AI fails is not the model — it is that nobody in the room was accountable for both the epidemiology and the engineering.
That is the whole thesis of the studio. Everything on this site is evidence for or against it, and I have tried to label which is which — including the projects in the incubator that are still validated only on synthetic data, and the ones still waiting on a funding decision.
If you are sitting on health data and trying to work out what is worth building, that is the conversation I want to have.