Clinical · plane-isolated
Clinical coworkers
A coder, nurse, pharmacist and scribe for every chart — all draft-until-signed.
Clinical coworkers are named AI agents that draft the documentation and clinical thinking around a chart: ICD-10/DRG codes, SBAR handoffs and care plans, medication second opinions, and the consultation note itself. Every output is a proposal a clinician must accept, edit or reject — no coworker can write to a chart hands-free, by design. They run on the Clinical data plane only, audited per decision with a full reasoning trace, and PHI can stay on-box.
Coder draft · ICD-10 + DRG
The clinical coworkers
A coworker for every clinical surface
Clinical isn't one job — it's the whole hospital. Each surface gets its own plane-isolated coworker that drafts the work and waits for a clinician to sign. Sixteen of them, and counting.
AI suggests: I50.0 + J18.9 (recommender-only)
4 / 4 dx coded
1 procedure coded
summary unsigned by MD
Coding coworker
draft-until-signedDrafts ICD-10 / ICD-9-CM / DRG on the signed chart.
Explore the surfaceauto-held: K+ > 5.5 — CDS rule
Ward nurse coworker
draft-until-signedDrafts SBAR, care plans and the e-MAR round.
Explore the surfaceCeftriaxone 2 g IV
Warfarin 3 mg
Paracetamol 500 mg
+1 more
Insulin RI sliding
Enoxaparin 60 mg
Vancomycin 1 g IV
Morphine PCA
Omeprazole 20 mg
Pharmacist coworker
draft-until-signedDrafts interaction checks, second-opinion & dispense verify.
Explore the surfacelistening — audio never leaves the browser
no raw transcript at rest · fail-open to regex
Ambient scribe
draft-until-signedDrafts the consultation note from ambient audio.
Explore the surface| NEWS2 aggregate ≥ 7 | CRITICAL | 2 today |
| Hyperkalemia K+ ≥ 5.5 | CRITICAL | 1 today |
| qSOFA ≥ 2 (sepsis screen) | WARNING | 3 today |
| NPO violation — oral order | WARNING | 0 today |
| Antibiotic timing 1h window | INFO | 4 today |
Every vitals write auto-fires CDS. Alerts surface via global FAB + drawer + inline badge. Rules are data rows — editable live, no code deploy.
CDS sentinel
draft-until-signedDrafts NEWS2 / MEWS / qSOFA / sepsis alerts on every observation.
Explore the surfaceOPD triage coworker
draft-until-signedDrafts an ESI triage level and the next station.
Explore the surfacered-flag: chest pain — AI deferred
ER coworker
draft-until-signedDrafts triage, bed placement and red-flag escalation.
Explore the surfaceLap chole · Dr. Wirote
next: TKA · Dr. Suda
block: General PM
set #A-114 reprocess
Block utilization · by surgeon service
Perioperative coworker
draft-until-signedDrafts counts, the surgical checklist and the op-note.
Explore the surfaceNewborn · ทารกแรกเกิด
✓HN issued
via global-sequence
AN linked
AN 69-002214
Apgar 1' / 5'
9 / 10
Labor-ward coworker
draft-until-signedDrafts partograph alerts and newborn registration.
Explore the surfaceICU flowsheet coworker
draft-until-signedDrafts flowsheet anomalies, trends and carry-forward.
Explore the surfaceLab coworker
draft-until-signedDrafts critical-value escalation and result triage.
Explore the surface04:12:33
01:21:08
00:42:51
158:07:44
Imaging & NM coworker
draft-until-signedDrafts report drafts and dose-safety checks.
Explore the surface▸ blood_lis_results present — register auto-unlocked
Blood-bank coworker
draft-until-signedDrafts compatibility, custody chain and reaction watch.
Explore the surfacecalc concentration = 4.0 mg/mL
BUD 24h @ 2–8°C · infuse ≥ 60 min
BCMA barcode
3/3 scans matched
Gravimetric verify
tolerance ±5% · 4.2 g / 4.0 g
Photo capture
step 2 of 4 captured
USP 800 hazardous
non-hazardous recipe
Double-check
awaiting 2nd pharmacist
Compounding coworker
draft-until-signedDrafts recipe math and incompatibility checks.
Explore the surfaceWard 5A
14 trays
Ward 5B
11 trays
ICU
6 trays
Nutrition coworker
draft-until-signedDrafts diet orders, NPO windows and meal rounds.
Explore the surfacePatient-360 assistant
draft-until-signedDrafts answers and actions across the whole chart.
Explore the surfaceWhat it senses
The signals it watches for you
It reads across the systems you already run — on the medOS event substrate — and surfaces the issue while there is still time to act.
Discharged-but-uncoded encounters
Watches encounter_journey_cache for newly discharged charts with no codes filed, then drafts ICD-10/ICD-9 and the DRG for a coder to sign.
Structured encounter summary
Reads the encounter's structured read model — diagnoses, orders, results, vitals — never free-text it has to guess at, so proposals are grounded in catalog IDs.
Live consultation transcript
The ambient scribe listens to the doctor-patient conversation locally (audio never leaves the browser) and turns the final transcript into chief complaint, HPI, narrated vitals and a SOAP note.
Active medication & allergy list
The pharmacist coworker reads the active orders, allergy list and CDS interaction rules to surface a second opinion before administration.
Vitals & observation writes
Every observation write fires CDS rules; the nurse coworker reads the resulting EWS trend and overdue acknowledgements to draft an SBAR escalation.
Shift & handoff boundaries
At end of shift it senses open alerts and carried-forward issues to draft the SBAR handoff for the oncoming nurse.
What it proposes
Drafted work, never an autonomous act
Each item lands in the Acknowledgement Inbox with its reasoning and a confidence score. Nothing is sent, charged or changed until a human accepts.
Drafted discharge coding
Primary I21.4 (NSTEMI) + secondary E11.9, drafted as a DRG with the supporting chart lines cited, queued to the coder to accept/edit.
SBAR escalation handoff
"S: 64M post-op day 2, HR 118, SpO2 91% on RA; B: …; A: possible PE; R: request urgent review" — drafted from the live EWS trend, sent to the charge nurse.
Pharmacist second opinion
Flags a potential warfarin + ceftriaxone interaction on the active order and proposes an INR-monitoring note — never blocks the order, only advises.
Ambient SOAP draft
From a 9-minute consult: chief complaint, HPI (onset 3d, severity 7/10), vitals BP 140/90, and a SOAP note — every field shown beside its transcript excerpt for per-field confirm.
The loop
Sense → propose → approve → execute
Sense
Reads the Clinical-plane read models only — discharged-uncoded encounters, the structured chart, a live consultation transcript, the active order/allergy list, or the EWS trend.
Propose
Drafts the code set, SBAR, second-opinion note or SOAP note as a proposal with a confidence score and the cited chart lines or transcript excerpts behind it. Nothing is written.
Approve
The draft lands in the Acknowledgement Inbox (or beside the consultation form) as an agent_proposal; a clinician accepts, edits or rejects each field — and that disposition is the training label.
Execute
On accept the agent calls the very same write path the clinician uses — the coding write, the note's Redux dispatch, the order form prefill — through the existing confirm-gate. No shadow write path ever touches a chart.
Capabilities
What it can do
Medical coder
Drafts ICD-10/ICD-9 and DRG for discharged encounters from the structured chart, citing the lines that support each code. The coder owns the code; the agent only proposes.
Nurse SBAR & care-plan
Composes SBAR handoffs and care-plan drafts from vitals trends, EWS scores and open alerts, so the escalation is written before the nurse picks up the pen.
Pharmacist second-opinion
Reads active orders against the allergy list and CDS interaction rules and offers an advisory note. Recommender-only — it can suggest, never gate or block an order.
Ambient clinical scribe
Continuous-listen consultation capture: audio stays in the browser, only the text transcript is processed, and the result pre-fills the same note form the doctor would fill by hand.
Catalog-grounded extraction
Differentials and drug names resolve through the SmartDiagnosis and order catalogs, not free-form generation — a hallucinated entity can't reach a proposal.
Per-field accept/edit/reject
Drafts arrive field-by-field beside their source, so a clinician confirms each one. The accept/edit/reject delta is the training signal, captured to the audit log — never to the chart.
On-box inference option
Extraction can run on Ollama on your own hardware, so transcripts and chart data never leave the building. Falls back to a local-regex path when no model is configured.
Not a black box
Why it is safe to run
Autonomy without guardrails is a liability in a hospital. These are the constraints that make this agent safe to put to work.
Draft-until-signed, always
No clinical write — code, note, order, vitals — is persisted until a clinician confirms it. Clinical writes can never run hands-free; it's enforced by construction, not policy.
Clinical plane only, DB-enforced
These coworkers run under a Postgres role scoped to the Clinical plane. Plane confinement is a database grant — a Growth or marketing agent literally has no grant to read a chart.
No hallucinated entities
The SeenMatches guard rejects any terminal proposal that references a code or drug not surfaced from the catalog during the run. The model identifies that a BP medication was mentioned; the catalog resolves the actual entity.
Audited per decision
Every proposal and disposition is logged to llm_audit_log under a named agent identity with the full RunnerStep reasoning trace; the accept/edit/reject delta is the label, never written to the chart.
Data plane
Runs exclusively on the Clinical data plane under a Postgres role with grants on clinical read models only; isolation is enforced at the database grant level, so a non-clinical agent cannot read a chart.
Operating characteristics
What changes when it runs
No autonomous clinical write
Every code, note and order is confirmed by a clinician before it persists
Always-on coding backlog
Drafts discharged-uncoded encounters overnight; your team works the exceptions
PHI can stay in the building
Extraction runs on Ollama on your hardware; air-gapped deployments supported
Granular human sign-off
Each drafted field is accepted, edited or rejected beside its source evidence
Works on your stack
Reads and writes where you already work
On accept, it calls the very same endpoints your staff use — no shadow write path, no second source of truth.
Questions
Frequently asked
Can a coworker change a patient's chart on its own?
No. By design, every clinical output is a proposal a clinician must accept, edit or reject. On accept it calls the same write path your staff use, through the existing confirm-gate — there is no shadow path that lets it write hands-free.
Where does the consultation audio go?
Nowhere. The ambient scribe uses the browser's local speech recognition; audio bytes never leave the device. Only the de-identified text transcript is processed, and it's cleared from memory after you apply or discard the draft — no raw transcript is stored.
Does PHI leave our network for the AI?
It doesn't have to. Inference can run on Ollama on your own hardware, including air-gapped deployments. When no model is configured the scribe falls back to a local-regex path, so the feature never breaks.
How does it learn?
From your clinicians. Every accept, edit or reject is captured as the training label in the audit log. The model sharpens against your team's real dispositions — it is never trained by writing to the chart.
Could a marketing or growth agent read a chart?
No. Agents are scoped to one data plane and that isolation is enforced at the database grant level. A Growth-plane agent's role has zero grants on clinical tables — it cannot read a chart even if it tried.
What exactly can the pharmacist coworker touch?
It reads active orders, the allergy list and CDS rules and drafts an advisory second opinion. It is recommender-only: it can flag an interaction and propose a monitoring note, but it can never gate, hold or block an order.
Put Clinical coworkers on your floor
See it draft real work against your own workflows — every action under human sign-off.