The model layer
Six models. One clinical record.
IASORA runs a named, versioned model layer rather than a single prompt behind a chat box. Each model does one job well on its own. Together they are what the agents are made of.
The six
Each one does a job you can name.
Speech becomes a note. The note becomes code. The code becomes a score, an answer, and a proposed action. Six models on one record, each doing one thing well enough to be measured on its own.
01 — Language to code
IASORA Lex
A note only becomes clinical data when something can name the code and defend it.
SNOMED CT · ICD-10 · LOINC · RxNorm · CPT
Conditions, medications, allergies, procedures, observations
Abbreviations, negation, disambiguation, specialty context, confidence thresholds
Scribe notes, referral letters, uploaded documents
02 — Speech to note
IASORA Scribe
The consultation is the richest record in medicine, and the one we most reliably lose.
In the room, on a video call, streaming or batch
50 languages, 72 regional variants
SOAP · progress note · discharge summary · consult letter
DocumentReference, conditions, medications, allergies, provenance
03 — Speech to speech
IASORA Voice
Most of what a practice does is a conversation. Voice makes those conversations write to the chart.
Real-time speech to speech, inbound and outbound
Enquiries · booking · reminders · recall
Against live availability
Communication, Appointment, Task
04 — Record to risk
IASORA Risk
A risk score nobody can take apart is a rumour.
Four weighted risk areas, 34 factors, a 0-100 composite, five tiers
NEWS2 · CHA₂DS₂-VASc · CKD-EPI 2021 · Charlson · Elixhauser · HHS-HCC · LACE · NHS eFrailty
Every point traces to a factor, a weight and its source
Deterministic baseline, accuracy and fairness gates, reversible in one step
05 — Question to answer
IASORA QL
The people who ask the hardest questions of a health record are rarely the people who can write SQL.
Count · list · trend · compare · breakdown · distribution
Analytical SQL, written against the FHIR schema
Every query validated before it runs, and workspace-scoped
Analytics Studio, the metrics builder, cohort definitions
06 — Evidence to judgement
IASORA Reason
Judgement is the part we refuse to automate. Reason prepares the thinking; the clinician still decides.
Deep, standard and fast
Differential reasoning, summarisation, structuring, decision support
The agent workforce
Your approval gate: nothing reaches the chart unapproved
The agent workforce
Each agent works next to a person.
An agent is a chain of these models pointed at one job. It prepares the work; the person it sits beside decides. Any model can also be licensed on its own.
Clinician
Documentation agent
Writes up the consultation and files it coded.
Patient
Inbound voice agent
Answers the call and books against live availability.
Practice manager
Outbound voice agent
Calls out for recall, reminders and results follow-up.
Care coordinator
Care-gap agent
Finds who is overdue, and reaches out to them.
Clinical director
Population health agent
Stratifies the panel and measures the outcome.
Whichever agent runs, the run is the same six steps
Every step above is one authorised, metered and logged call against your workspace: which model, which user, what it was asked, what it returned.
How they run
Every call is authorised, metered and logged.
The model layer sits behind one gateway. That is what makes it reviewable: there is a single place where access, cost, residency and the audit trail are decided.
On or off, per workspace
Each model is switched on for a workspace at a tier, or left off entirely. A clinic that wants the scribe and nothing else gets exactly that.
Metered to the action
Every call is priced and counted where it happened, so the cost of an AI action is a number you can put next to the work it did.
Logged, every request
Which model, which workspace, which user, what it was asked and what it returned. The audit trail is written whether or not anyone is looking.
Inference stays in region
Kept inside your jurisdiction, on private cloud or on-premise including air-gapped, for health systems that cannot send data out.
Bring your own keys
Run the layer on your own model credentials where policy requires it, without giving up the routing, metering and audit around them.
Versioned, and reversible
Models carry versions. Weights, thresholds and rules change under an audit diff, and a change that does not hold up is rolled back in one step.
Where the models show up in the product
Scribe and Voice in the consultation and at the front desk, Lex on every document, Risk across the panel, QL in Analytics Studio, and Reason inside every agent.
Get started
Run one model, or the whole layer.
Use the scribe on your own consultations, or talk to us about the model layer inside your own estate.