One pipeline.
One patient model.
Everything the pipeline knows about a person lives in one timestamped record, written in the terminology medicine already agrees on.
The half of your health a clinic never sees.
Most of a person's health happens between appointments, and almost none of it gets written down. Our patient portal collects it without asking anyone to keep a diary: the signals your devices already produce arrive on their own, and anything they cannot see takes one reply.
of a person's health is influenced by non-medical factors, excluding genetics. Those factors live in daily life: lifestyle, stress and behaviour, not the few minutes of a consultation.
-
Connect
Link the wearables and health apps you already use. Cloud services sync server-side: the vendor tells us the moment a record lands and we pull only that record. Phone-native sources push straight from the device. Either path runs in the background.
-
Normalise
Every source records data its own way. The portal translates each record into one shared format, removes duplicates across devices, and timestamps every reading in your local time. One consistent stream instead of a dozen feeds that do not line up.
-
Converse
The portal watches the normalised stream and messages you first when something shifts, then answers when you text back. Reply in plain language; that is the logging. No new app, no forms, no streaks.
Every device, one format.
Sleep, movement, meals and mindfulness are recorded in one shared format, so a figure means the same thing whichever device reported it. Two people with different watches get comparable numbers.
- Sleep Duration, efficiency, deep and REM
- Recovery & HRV Readiness, recovery score, resting heart rate
- Activity Steps, movement, daily exertion
- Nutrition Meals and substances, from a normal text
- Mindfulness Meditation and breathing sessions, minutes and streaks
- Context Screen time, location and daily rhythm
Health records can be read into the portal but never acted on by it. The portal is a general-wellness product under the FD&C Act §520(o)(1)(B) exclusion, not a medical device: it gives wellness guidance, not diagnosis.
One timestamped model of a person.
A person's health is scattered across devices, labs and memory. We hold it as one record: every observation written in the same clinical vocabulary and stamped with when it was true.
Because every entry carries its date, the record answers questions a snapshot cannot: what a reading was, how it has moved, and whether the direction matters. Iron studies rising. A resting heart rate creeping up over months.
Acute readings arrive the same way. Our own devices (a recording stethoscope, a blood-pressure monitor, an otoscope) stream straight into the core: a closed hardware-to-software link, with no manual entry and no third-party integration in between.
- 01 Everyday-life signals from the patient portal
- 02 Acute vitals, streamed from our own devices
- 03 Lab results & records
- 04 Clinical notes & history
- 05 What a person tells us, in their own words
Resolves to One timestamped patient state
A loop, not a pipeline.
Most health AI answers whatever it is asked, once. A clinician does something else: they form a view, work out what would change it, and go and find that out. Auracare runs the same way, and keeps going round until asking again would not be worth the delay.
-
Input
One timeline: everyday-life signals from the patient portal, clinical history, and vitals captured live in the room, all on a single timestamped record.
-
Encoding
Each observation is matched to its concept in the clinical ontology and stamped with when it was true. This is the step that turns loose readings into a record the core can reason over.
-
Thinking
The core weighs the evidence across the knowledge graph and returns a spread of what is likely, rather than committing to one answer.
-
Thesis
That spread is checked against population data, expressed as risk, and set out as a ranked differential with its sources attached. It is provisional, and it changes when the evidence does.
-
Value-of-information
The loop weighs whether another question would change the answer enough to justify the delay. If it would, it picks the single most useful one and starts again. If not, it stops here.
-
Medical outcome
A referral, a prescription, further testing or a lifestyle plan, each checked against safety and against what the clinician is permitted to do where they practise.
Round again, until another question would not change the answer
Stage five is the only way out of the loop. Nothing reaches a patient before it, and everything that follows it is checked by the safety layer on the way past.
Two kinds of intelligence, checking each other.
Learned models are fluent and occasionally confident about things that are false. Rule-based systems are reliable and cannot handle anything they were not told about. We run both and keep a marked line between them, because the question a regulator asks is which side of that line a given decision came from, and whether you can reconstruct it.
Proposes
Learned
Adaptive, advisory
- A Heterogeneous Graph Transformer: a neural network that learns patterns across a patient’s linked clinical data
- Bends the generic textbook weights toward this patient’s comorbidities and trajectory, then ranks the shortlist
- Advisory by construction: it proposes and personalises, but never makes an un-gated decision
Disposes
Symbolic
Auditable, authoritative
- The clinical ontology (SNOMED CT, ICD-11, LOINC, HPO): curated, deterministic, every edge inspectable
- Runs the red-flag and contraindication screens, and reads diagnoses straight off named graph edges
- Authoritative: it holds the veto over anything the learned side proposes
Every conclusion traces back to the specific rule or relationship that produced it, and every step is logged in an order you can replay.
One gate on every stage.
One check at the end would mean everything before it is allowed to go wrong, provided the last step catches it. The checks run at every stage instead, and take a different form at each.
The outermost gate is a person. Auracare is built to be used alongside a clinician, never to replace one. The clinician stays accountable for every decision, and their judgement always overrides the model.
- Input · Encoding
Encoding-confidence gate
A signal that cannot be matched to the right concept with enough confidence is flagged rather than quietly accepted. One mis-linked observation would corrupt everything downstream of it.
- Thinking
Audit log of decisions
Every reasoning step is recorded with its inputs and provenance, so any conclusion can be reconstructed and replayed after the fact.
- Thesis
Red-flag screen
A hard, authoritative screen over the differential that can escalate or veto regardless of what the learned side proposed.
- Value-of-information
Jurisdiction guard
The loop only ever proposes actions permitted where the patient is: prescribing authority and what is possible in primary care versus referral.
- Prescribing
Contraindication check
The last gate before any terminal action: interactions, allergies and pharmacogenomic contraindications, checked against the medicine.
- On live output
Clinical assurance sampling
Clinicians grade samples of real output against a harm scale, with sign-off before each new deployment and random re-review after it.
Data residency is jurisdiction-based. The reasoning core is designed to run inside our own cloud tenant in each market's region, so consented health data stays within infrastructure we control. An architectural commitment for an engine still in development, whose regulatory pathway is under continuous review.
Built on the terminology medicine already uses.
Both products read from the same graph of clinical concepts and the relationships between them. None of it is scraped from the open web: it is mapped to the standards clinicians, regulators and health systems already use. It is also the one part of the system that is finished, and you can go and look at it now.
- SNOMED CT The anchor
- ICD-11 Diagnoses
- HPO Phenotypes
- LOINC Lab codes
Every answer traceable to a named source.