Individualized AI models, methods, software and wearables for medical research, therapeutics and care.
Doctors treat one patient at a time. But almost every clinical model treats patients as group members. We close that gap: we build software and wearables that collect data from the individual patient, and instruments that can distinguish treatment from biological response.
For drug discovery, design and development · procedural medicine · clinical trials · patient care
Digital twins and related technologies for drug discovery, design and development are built on averages — a reference cohort supplies the answer and the individual is scored against it. AI as a Biological Primitive (ABP) inverts that. It assigns an AI to each relevant component of a patient's system and learns its biology from that system's own responses, then interrogates the result in reverse: specify the state you want, and it returns the interventions that patient's record actually supports.
Trained on
ONLY 24.5% of the time did the population average model agree with ABP, which was based on individual biology. Personalising a population model is not the same as modelling the person.
Which drug, at what dose, on what schedule, in whom — and where the risk sits. ABP tests each option against a model of the individual patient, and refuses to recommend anything that patient's own history does not support. That refusal is the point: a population model will answer any question you ask it.
Open the ABP demo →During a procedure the question is not what usually happens. It is whether this patient's response has changed, and whether the view supports the call being made. Deferra separates a poor picture from a departed course from a changed response — three situations that look identical on a monitor and call for opposite actions.
Decides when an AI system should assist, hold back, or say nothing as a procedure evolves — and when the view cannot support the call, tells the surgeon what to correct rather than simply going quiet. Its default state is silence.
Open the surgical demo →The same logic under anesthesia. Every dose changes how the patient responds. Deferra watches for a change in the response itself, not for a number crossing a line.
Open the anesthesiology demo →Builds a model of one biological system from that system's own longitudinal history, then interrogates it in reverse to identify supported intervention directions and levels for that patient.
Open the intervention demo →Methods that compare a patient against a model-derived expectation can produce large groups of apparent responders and apparently harmed patients even when the treatment did nothing at all. CTF tests whether those groups are real by running the identical procedure on patients who received placebo — who cannot have responded to a drug they never got.
Tests whether an apparent responder subgroup survives its own negative control — before a trial is designed around it. Stress-tests what a design can actually falsify, and separates a population-level result from stronger claims about individuals, subgroups or mechanism.
Open the CTF demo →These instruments sit upstream of modelling. They determine what a measurement can support, what data to acquire next, which turning points must be remembered, and when a system should act — or decline to.
Say what you need to see, and OptiCeil returns the imaging setup that will show it — and the one thing standing in the way when it cannot. It reads the recording in front of it rather than a population average.
Open the OptiCeil demo →Tells you, before anything is built, whether the data contain the distinction you are asking for — and if not, what would have to be measured for them to.
Open the NPIS demo →Marks the moments where a patient's course actually turned — the points you would need to reconstruct the path rather than average it away — and flags when new data breaks what was stored. TotemicAI-Specified uses a declared structural scoring policy; TotemicAI-Learned fits the weighting policy from training data before compiling the landmark memory. Current public AML validation is for the Specified mode; Learned is evaluated separately under patient-held-out validation.
Open the TotemicAI demo → Read the TotemicAI technical report →Determines what data to acquire next, in whom, and at what point further measurement stops earning its cost. In one study, measuring 30% of patients captured 71% of the value of measuring all of them.
Open the Acquire demo →Takes a field's published claims, reruns them on one common dataset under one protocol, and reports which reproduce and which competing explanations the data can actually tell apart. Applied to the whole Alzheimer disease literature.
Open the AIO demo →Returns a decision rather than a score: use it, keep going, acquire selectively, change the measurement, add a modality, reacquire, or abstain.
Open the TopolAI demo →Modelling one patient requires enough of that patient's own history to learn from. Continuous sensing is what makes it possible — and what you measure determines what can be recovered. On the same wearable recordings, adding motion sensing improved the result where changing the model had not: what you measure matters more than which model reads it.
Wearables, continuous glucose monitoring, photoplethysmography and physiologic waveform capture, evaluated for what they can actually support rather than for how much they collect. The instruments above determine which signal is worth adding before it is deployed.
See the measurement instruments →
CMH Scientific is where the instruments are built and validated. Each one has an interactive demonstration that runs in a browser with nothing to install, alongside the technical papers and the studies behind it. Every instrument states what it has been shown to do — and what it has not.
Every instrument has an interactive demonstration that runs in a browser with nothing installed. Production access is licensed and product-scoped; academic, nonprofit and commercial use each require verified identity, institution and intended use. Protected engine logic is not distributed with the public demonstrations.