Making human physiology computable.
Lumos is building a foundation model of human physiology — a continuous-time, multimodal model that learns how the whole body actually works from raw physiological signals. The missing primitive beneath modern medicine and drug development.
The Divergence
Medicine never had its computational revolution.
Supply chains and telecommunications were rebuilt by computation. They learned to predict, to scale, to drive error toward zero. Medicine never had that revolution.
In 1952, a critically ill patient was watched at the bedside as their chest rose and fell. In 2026, that hasn't fundamentally changed — there is simply a monitor in the room now. The tools got better. The outcomes barely moved.
Eroom's Law
Drug development keeps getting harder, not easier.
Every other technology gets cheaper and faster over time. Drug development does the opposite — so reliably the industry named the curve Eroom: Moore's Law, in reverse.
The Signal
The body produces millions of signals a day. Medicine keeps a handful.
Data points from a single bedside channel each day, reduced to one number.
A continuous ECG captures 250 readings every second — the electrical state of every chamber, the autonomic response to every drug. The monitor reduces all of it to one number, checks a threshold, and discards the rest.
The same happens on every channel. A full arterial waveform becomes “BP 120/80.” The oxygen curve that encodes respiratory effort and perfusion becomes “SpO₂ 97%.” The signal was never noise — there has just never been a model that could read all of it at once.
Why we're building this
Her voice was changing. Nobody was listening.
My grandmother had Parkinson's. The depression that came with it — a known, treatable comorbidity — went unrecognized until it took her life. I was eight when I learned how she died.
Years later, I read that vocal biomarkers — micro-tremors, shifts in pitch and phonation — can flag Parkinson's progression and depression from a ten-second recording. Every phone call she made was data. Every conversation was signal.
She used to call me Zayka. Little rabbit. The signs were there the whole time — in her voice, her movement, her sleep. Nobody was watching for them.
Everyone knows someone the system wasn't watching for. That is the future we are trying to make impossible.
What we're building
Stop compressing. Learn from the raw signal.
Instead of training on the categories we already record, we train on the raw, full-resolution stream the body produces — and let the model learn how the body actually changes and responds: to disease, to drugs, to intervention, moment to moment, organ by organ.
Every sensing modality gets its own encoder. A shared backbone maps them into one continuous, evolving representation of a patient's state. New sensors plug in without retraining, and the model keeps learning.
We're early, and the science is genuinely hard. But the signal is real — and we're building to hear all of it.
Collaborate
Work with us.
Lumos is early. We're building our foundation model of human physiology alongside the people closest to the signal — clinicians and the institutions that hold continuous physiological data.
team@projectlumos.orgFor clinicians
Pressure-test the science, shape what we build, and tell us where the model is wrong. We want the people closest to the bedside in the loop early.
For hospitals & data partners
Partner with us on continuous physiological data. The richer and more continuous the signal, the more the model can learn about how the body actually behaves.