A loop, not a pipeline.
Most health AI runs a fixed pipeline: data in, answer out. We are building something that behaves more like a clinician — a reasoning loop that reasons, then asks the next-best question, then reasons again, always grounded in the knowledge medicine already agrees on.
Five stages that feed back into themselves.
A pipeline ends when it produces an answer. Our loop treats every answer as the start of a better question — and a translucent safety layer wraps all five stages, end to end.
-
Input
Signals arrive — wearables, labs, notes, a question asked in plain language.
-
Encoding
Each signal is mapped into one shared clinical vocabulary, with a confidence score attached.
-
Reasoning
The neuro-symbolic core weighs the evidence over the knowledge graph to narrow what is likely.
-
Output
A grounded, traceable response — every claim linked back to a named source.
-
Value-of-information
The loop decides the next-best question — the one that removes the most uncertainty — and asks it.
Safety runs across all five. No stage is left to reason unsupervised — the same guardrails apply from the first signal to the next question asked.
Two kinds of intelligence, checking each other.
Learned models are powerful and fluent, but they can be confidently wrong. Symbolic systems are rigorous but rigid. We pair them — and give the auditable side the final say.
Auditable, authoritative
- Grounded in SNOMED CT, ICD-11 and the terminologies medicine already agrees on
- Deterministic and inspectable — you can follow every step
- Authoritative: it holds the veto
Adaptive, advisory
- Reads across messy, multi-modal signals to spot patterns
- Adapts as more consented data flows in
- Advisory: it proposes, it never decides alone
A learned model proposes, an auditable symbolic layer disposes.
One timestamped model of a person.
A person's health does not live in one app. It is scattered across devices, labs and memory. The digital twin pulls those sources into a single, timestamped state — so the reasoning loop always works from one coherent picture, not a dozen partial ones.
Every source is encoded into the same vocabulary and stamped with when it was true, so the twin can answer not just what a reading was but how it has changed, and whether that matters now.
- Wearables & continuous sensors
- Lab results & records
- Clinical notes & history
- What a person tells us, in their own words
- One timestamped patient state
Grounded in what medicine already knows.
The reasoning is anchored to a live graph of clinical concepts and the relationships between them — the part of the system you can explore for yourself today.
Traceable to a named source.
Our knowledge is not scraped from the open web. It is mapped to the standards clinicians, regulators and health systems already trust.
- SNOMED CT Clinical terms
- ICD-11 Diagnoses
- HPO Phenotypes
- LOINC Lab codes
- NICE UK guidance
- MHRA UK regulation
- dm+d Medicines
Every answer traceable to a named source.
Built to be checked.
Trust in a health system is earned in the details. These are not features bolted on at the end — they are how the loop is built.
Encoding-confidence gates
A signal that cannot be mapped with enough confidence is held back, not guessed at.
A full audit log
Every input, inference and source is recorded, so any answer can be traced after the fact.
Red-flag routing
Anything that reads as urgent is routed to the right emergency or clinical service, immediately.
Jurisdiction awareness
Guidance and routing adapt to where a person actually is — the standards and services differ by country.
On data residency: the reasoning core is designed to run inside our own cloud tenant, so consented health data stays within infrastructure we control. This is an architectural commitment for the agentic engine — which remains in development.