How it works
An overview of the engine behind SomaTapestry: what the model is, how it becomes personal, and how it is kept accurate as it grows.
A graph-based model of metabolism
SomaTapestry represents human physiology as a graph. The organs, tissues and circulating substances that hold state are connected by the processes that move mass and signal between them, and medications, supplements, meals and lifestyle inputs act on that shared structure rather than on separate models. This is what allows a glucose forecast to account for a medication taken that morning, the previous night's sleep and the meal just logged in a single calculation.
The model runs forward in time and produces a predicted trajectory with an uncertainty band for every signal. Any point on a curve can be traced back to the mechanisms that produced it, which is the basis for the explanations shown in the apps.
Coverage spans the metabolic core, major hormone axes, a growing library of medications and supplements, and lifestyle drivers such as exercise, fasting, sleep and travel across time zones.
Calibrated to one person
The model starts from a reference profile and is adjusted in stages. Sex, age, body composition, conditions and known genetic variants set the starting point. Connected devices then supply measurements: when a reading exists it anchors the model, and between readings the model fills the gap.
After each meal, the engine compares its forecast with the measured glucose response and refines the small set of parameters that the data can inform, such as insulin sensitivity and glucose absorption. New estimates are blended with previous ones so that a single unusual day does not distort the model. Users can see which parameters are well determined from their own data and which remain population estimates.
Evidence and review
Every change to the model carries a decision record: the claim being made, the primary source supporting it, the parameters affected, and the test that would show it to be wrong. The validation criteria are fixed before a change is scored.
A change is admitted only when it passes the existing validation cases and is approved by a reviewer. Deferred and rejected proposals remain on record with the reason, so the history of the model is auditable.
How the model grows
Expanding a physiological model by hand is slow. SomaTapestry uses AI agents to search the literature, extract candidate parameters and mechanisms with their sources, run validation against each proposal, and suggest calibration adjustments. The agents prepare proposals; they do not edit the model. Every proposal passes through the same evidence review as a change made by a person, and agent activity is rate-limited and logged.
From a meal photo to a forecast
A photo is analysed to identify the items on the plate and estimate portions. Nutritional composition is then calculated from a curated food catalogue rather than estimated by the vision model, and the user confirms any item with low confidence before the forecast runs. The resulting meal is passed to the engine, which forecasts the glucose response from the user's current state. When the sensor later records the actual response, the meal becomes a calibration point.
Meals can also be logged as text, from favourites or from recipes.