Home/Standing Watch/Datasheet
AI MedAgent Standing Watch
One reasoning layer from the ICU to the kitchen table

What it is

Between the rooms, nobody is watching. A patient leaves the ICU, then the ward, then the hospital, and every hand-off drops something. Standing Watch is AI MedAgent stretched across the whole stay and the first 30 days at home: the same reasoning layer in every room, a mediator on every finding so it reaches a person and gets an answer, and an encrypted bridge to the person's own phone, where the consumer app, Ami, takes over. The hospital runs the watch through day 30 inside its own security boundary. After that, it is the person's own.

AI MEDAGENT STANDING WATCH · ONE WATCH, FIVE ROOMS AI MEDAGENT · THE HOSPITAL RUNS THE WATCH AMI · ON THE PERSON'S PHONE ICUCLOCK: MINUTES Reads pump, monitor,orders, labs, the labelHears: ICU nurse,pharmacist, intensivistLive reasoning WardCLOCK: HOURS Reads the reconciledlist, new orders, labsHears: floor pharmacist,hospitalistWalkthrough DischargeCLOCK: ONCE Builds the home packetand the pairing codeHears: hospitalist; packetto patient and familyEncrypted in browser Home, 1 to 30CLOCK: DAILY Health app list,packet, public lensHears: the person;questions to the nurseAmi, take one After day 30CLOCK: DAILY The person's ownrecords, public lensHears: the persononlyAmi, take one SMART HealthLink key in linkblind server Reasoning layer A large language model over the local feeds and the free public data marts (openFDA, DailyMed, RxNorm, the UMLS, MedlinePlus), after the Rosetta layer normalizes them The mediator, on every finding recipient role and chain · criticality and its reason · reminder timer · escalation timer · respond, defer, reject or nothing · an audit entry per step Carried in HL7 FHIR R4 Task, CommunicationRequest, DetectedIssue and AuditEvent; IHE Alert Communication Management for device-originated alerts One audit trail: ICU to day 30, kept by the hospital inside its own fence. After day 30, nothing leaves the phone but a question the person chooses to send.
Figure 1. Navy dashed: AI MedAgent, run by the hospital. Light blue dashed: Ami, on the person's phone. The clock slows, the recipient changes, the audit trail never breaks.

The method, in four steps

StepWhat happens
1. AcquirePump (IHE PCD-01), monitor, orders, labs and allergies (HL7 FHIR); at home, the Health app list, wearables and the hospital's discharge packet.
2. NormalizeThe Rosetta layer. Whatever the EHR, whatever the pump, one meaning comes out: RxNorm for drugs, LOINC for labs, the ISO/IEEE 11073 nomenclature for devices, the UMLS over all of it. The same job IHE's Rosetta Terminology Mapping does for pumps, done for the whole encounter.
3. ReasonA large language model, continuous, over the local feeds and the public data marts. Proposes act, hold or escalate, inside bounds the clinicians set, and says why and what it could not check.
4. Connect to a humanThe mediator carries every finding to a named role and keeps it there until someone answers. In the hospital, the nurse and the pharmacist decide. At home, the person reads one sentence and may send one question.

The mediator, field by field

FieldCarried in (HL7 FHIR R4)Demonstrated
The finding and its evidenceDetectedIssue: code, severity, detail, evidenceYes
Who hears it, and who is nextTask.owner (PractitionerRole); CommunicationRequest.recipient; the chain in Task.inputYes
Criticality, and whyTask.priority (routine, urgent, stat); the reason in Task.input; from the CDS Hooks card indicator when the live reasoning writes itYes
Reminder and escalation timersTask.restriction.period.end; minutes in Task.inputYes, on a simulated clock
Respond, defer, reject, or nothingTask.status: completed, on-hold, rejected; Task.statusReason; Task.noteYes
The trailAuditEvent, one per step; downloadable as a FHIR BundleYes
Device-originated alertsIHE Alert Communication Management, PCD-04 to PCD-07Mapped, not yet wired
Delivery channelCDS Hooks card first; secure chat where an API existsNamed, not yet wired

The bridge, discharge to home

What runs on aimedagent.net

DemonstrationAddressWhat is real
Reasoning Demoaimedagent.net/reasoningLive reasoning on a synthetic critically ill patient, with the federal evidence it read
ICU watch boardaimedagent.net/watch-icu.htmlThe mediator on five synthetic beds; any bed's chart can be sent to the live reasoning
One patient, five roomsaimedagent.net/watch-journey.htmlThe mediator and one audit trail across the continuum; scenario findings
The bridgeaimedagent.net/watch-bridge.htmlReal SMART Health Link construction and AES-256-GCM encryption in the browser; Ami side simulated
AmiiPhone, take oneBuilt and running on the developer's own phone; not yet distributed

Standards and sources

Status and limits

AI MedAgent is covered by multiple pending U.S. patent applications, in which AI reasoning over a patient's local data together with public data marts is the central claim. Standing Watch is a research demonstration on synthetic patients, built and funded by one person. It is not a medical device, is not FDA cleared, and is not for clinical use. A hospital that ran the watch for real would set its own recipients, timers and escalation chains, run it inside its own security boundary, and validate it before anyone relied on it.

Working with me. I work with device makers, health systems and health IT companies one at a time: licensing the method, a pilot partner for the first real watch, or advisory work.

For additional information, contact Daniel Pettus at dan@aimedagent.net or through aimedagent.net/contact.
AI MedAgent Standing Watch · aimedagent.net · Daniel C. Pettus, October 2026
Patent pending. Synthetic patients only. Not a medical device.