One idea about how a machine should see. Its consequence was a model small enough to fit in the places medicine forgot.
Chest X-rays are among the most common and useful tests in medicine. Across much of the world, there are simply not enough radiologists to interpret them.
The AI tools that could close that gap were built for the places that need them least. They assume fast internet, cloud servers, and hardware that under-resourced clinics do not have. A tool that requires a data centre cannot help a clinic with unreliable power.
So the technology that could serve the underserved was built, by default, in a form they cannot use. That is not a failure of intent. It is a failure of architecture.
It would be easy to tell a more flattering story. This is the true one.
GeoRadX did not begin with a patient. It began with a question about how machines see.
Before there was any thought of medicine, there was an engine — a recursive architecture built to reason about space rather than symbols, and validated on abstract reasoning problems that have little to do with health. It was pure research, driven by the kind of intellectual restlessness that neither recognizes boundaries nor rules.
Then came an X-ray. And with it, a realisation that seems obvious once said aloud: a pathology has a shape. An anomaly has a form, a place, a spatial signature. You do not need to be a radiologist to see that something in an image is wrong — that the geometry of the thing departs from the geometry of health — even if you cannot name the disease.
Modern medical AI mostly learns from labels: millions of images, each tagged with a name, from which a very large model infers a correlation. But that is not what a clinician does when they look at a film, and it is not what the eye does. The eye sees structure.
Don't teach machines labels. Teach them geometry.
That single reframing is the origin of everything else. And its consequence — the reason this page exists — was entirely unplanned.
We did not design for rural clinics and then look for a technology. We followed an idea, and the idea led here.
This is the part worth being precise about. The accessibility was not added later. It fell out of the science. A model built on geometry is small; a small model needs no infrastructure; a model that needs no infrastructure can be handed to a clinic with unreliable electricity and a decade-old desktop.
Somewhere in that chain, the project stopped being an experiment about machine perception and became an obligation. Once you understand that the thing on your screen could give a first read to a patient who would otherwise get none — curiosity is no longer a sufficient reason to keep going, and it is no longer a sufficient reason to stop.
GeoRadX runs on DRF-MWU — a recursive field engine that resolves an image by converging on its spatial structure, rather than classifying it against a library of labels.
A theory of how a reasoning system holds a stable reference to what is actually there while its own interpretation moves. It is what gives the engine a fixed idea of "healthy" to reason against.
The architectural principles that govern how such a system may be built and extended without breaking its own geometry. It is why the engine stays small and stays coherent.
The ambition isn't only to put a diagnostic engine in more clinics. It's to change what a clinic in an under-served place is capable of.
Most health technology offered to these settings is a centralised platform — something you connect to, that holds your data and hands back a result. GeoRadX is built the other way around: a precision instrument that lives with the clinician, works whether or not there's a connection, and does three things a platform alone cannot.
It monitors — following a patient's disease across visits, region by region, with numbers rather than memory. It diagnoses — a fast, grounded second read where no radiologist is in the room. And it enables research — turning the ordinary work of care into a de-identified evidence base a health programme can actually study. That is how we intend the technology to advance the future of both clinical work and patient care: not by centralising it, but by making the point of care itself more capable.
I have spent most of my life being made capable by tools other people built. Someone I will never meet designed the machine I work on. Someone else built the network that put the world's knowledge within reach of a curious person in Nigeria. Others built the AI systems that let one researcher, working part-time, do what used to require a team.
I did not earn any of that. It was simply there, made by people who decided to build something useful and then let strangers use it.
At some point it stopped being enough to be someone who uses useful things. I wanted to be someone who makes them.
GeoRadX began as a question about machine perception, and I have been honest on this page about that. But an idea does not stay where you leave it. Once I understood that a model built on geometry would be small enough to run on a clinic's old computer, with no internet, in places where a radiologist may be a day's travel away — the work acquired a debt it had not asked for.
I do not know yet whether GeoRadX will matter. It is not approved. It has never been used on a real patient. It may fail. But it will not fail because someone cut a corner to get it out faster, or because a clinic was told it could do something it cannot.
If it reaches a clinic, it will be because it earned the right to be there.
The clearest way to judge this is to look at the reading itself — what it flags, and what it shows you about why.