I have liked working since I was a child. Not being busy, which is a different thing, but building something and watching it start to hold weight.
Formulas before products
I was nine or ten when I began caring about what mathematics and physics actually do once they leave the page. Not the exercises. The other part, where a formula turns out to describe something real, a force, a rate, a thing that happens whether or not anybody writes it down. I liked inventing my own. Most of them were wrong and a few of them were useful, and either way the habit stuck. Look at a problem in the world, assume there is a formulation of it somewhere, and go and find it.
At twelve I found machine learning, and that settled it. It was the first thing I had seen that let a formula correct itself against evidence instead of sitting still on the page waiting to be right. This is my thirteenth year of working with it, which means I have now spent more of my life doing this than not doing it.
I never really stopped. Kantesti is the same instinct with better tools and a much larger dataset.
2019, a sofa, and a very small model
The foundations were laid in 2019. I was sitting on a sofa and asked myself a fairly small question. What if I trained a very small language model, a few million parameters, purely to write one or two sentences summarising a blood test? Not a diagnosis. Not a product. A summary, in a sentence, that a person could actually read.
I started that day. The first version was a natural language model of four and a half million parameters, and it did roughly what I hoped, which is to say badly, but recognisably. It was enough to see the shape of the thing.
A few million parameters, one or two sentences, and a page of numbers that had never once been written for the person holding it.
Six years
Days chased days. At the end of the sixth year that four and a half million parameter experiment runs at 2.78 trillion parameters, and it no longer only summarises. It interprets blood tests and DNA tests, recommends supplements, produces health reports and advises on diet. As far as I know it is the largest engine in the world doing this particular work, and it began as one question asked on a sofa by somebody with no reason to be asking it.
The part I did not expect was the languages. It became obvious quite early that a translated interface is not the same thing as a local one, because a Hungarian laboratory report is not an English one with different words in it. So each country got built properly, against its own reports, its own reference ranges and its own units. That decision is why there are fourteen properties instead of one.
What it is actually for
A list of fourteen companies does not explain itself, so I should say plainly what they are in service of. The aim is to raise the quality of health for as many people as possible, everywhere, and not only for the people who can afford to have somebody sit down and explain their results to them.
A blood test is the cheapest and most widely available window into a human body that medicine has. Hundreds of millions are produced every year, and a very large share of them are read by nobody except the machine that files them. Closing the distance between a test being taken and the person understanding what it says is, as far as I can work out, the largest available improvement to human health at the lowest possible cost. Everything above this paragraph is a means to that, including the parts that look like business.
Leaving Milan
In 2021 I was reading artificial intelligence at the University of Milan, and I left. Not because it was going badly. It was going well, which is the only reason the decision was difficult. I had simply worked out that academia was not the right road for me. I wanted the version of this problem that has real users attached to it and a bill at the end of the month, so I carried on through trade and innovation instead. I have not once needed to revisit that.
Why there is no investor on this page
People assume bootstrapping was a principle. It was not. It was a consequence. When you are one person and the work is to ship in seventy-five languages, the slowest possible path is the one where you first spend six months explaining the idea to somebody whose job is to be sceptical of it.
So there was no round, and there is no board, and there is no runway, because the fourteen properties pay for the fourteen properties. It has cost me things. There is nobody to hand the hard part to, and the hard part arrives most weeks. But it also means that nothing on this site had to be approved by anybody before it existed.
Eleven countries, and a train
I travel constantly. I visit our offices across eleven countries, and most of what gets built starts on one of those trips, usually in a conversation with somebody who reads laboratory reports for a living, and occasionally on the train afterwards. I have never been good at holding on to an idea for long. I would rather build it badly on a Tuesday and find out.
That is most of it. All fourteen are set out on the companies page, the public record sits on the home page, and anything you want to ask goes to the address on the contact page.