Technical overview

A standards-native record, two equally useful views, a reasoning layer, and a trust plane.

Guardian Angel is built from commodity, interoperable components. Its novelty is not in any single layer but in their composition: a patient-controlled longitudinal record at the center, symmetric clinician and patient views, an agent runtime that reasons over the record, and a governance plane that mediates every read and every action.

Detailed architecture INGEST · FHIR / SMART / device protocols EHR connectors Wearables / phones Labs & imaging Pharmacy Payer/claims Genomic RECORD · patient-controlled longitudinal Guardian Angel Record FHIR-native · multi-modal · provenance-tagged · consent-bound REASONING · agent runtime Trajectory engine cross-visit patterns Summarizer role-aware views Guideline grounding cited recommendations Workflow orchestrator tools · alerts · handoff VIEWS · symmetric clinician + patient Patient view (+ family delegate) conversation · preferences · reminders Clinician view (+ care team) summaries · trajectories · escalation Trust plane · identity · consent · audit · escalation · provenance
Four planes — ingest, record, reasoning, views — with a continuous trust layer underneath.

Data model

FHIR-native, multi-modal, and provenance-aware by construction.

Clinical

Standard FHIR resources

Patient, Encounter, Observation, Condition, MedicationStatement, AllergyIntolerance, Procedure, DiagnosticReport. Sourced from EHR connectors, normalized at ingest, never overwritten on conflict.

Devices

Continuous observations

Wearable and phone-derived streams represented as time-series Observations with explicit confidence and sampling metadata. Aggregated into clinically meaningful summaries by the reasoning layer rather than dumped into the clinician view.

Imaging

DICOM references with image-derived signals

Imaging studies referenced rather than copied, with structured findings, radiologist impressions, and AI-derived measurements stored alongside the pointer.

Genomic

Variant-level and report-level genomic data

Standard genomic FHIR profiles plus pointers to source variant files. Patient-actionable findings surfaced explicitly; secondary findings flagged for consent.

Lifestyle

Patient-entered context

Diet, exercise, sleep, mood, exposures. First-class but provenance-distinguished from clinical observations. Used by the reasoning layer; surfaced to clinicians only when relevant.

Preferences

Care and communication preferences

Goals of care, decision-making style, language, channels, alert thresholds. Stored in the record so they travel with the patient across institutions and countries.

Agent topology

One reasoning layer with four explicit responsibilities.

The agent runtime is not a single monolithic prompt. It is a small set of specialized reasoners coordinated by an orchestrator, each grounded in the record and operating under the trust plane.

Trajectory

Detects cross-visit patterns

Identifies drifts, regressions, and inflections in measurements and conditions across the full longitudinal record. Outputs are framed as hypotheses with confidence, not diagnoses.

Summarizer

Role-aware rendering

Produces patient-facing and clinician-facing renderings of the same underlying record. The two views are deliberately commensurable: what one party sees is auditable against what the other sees.

Grounding

Guideline-cited recommendations

Recommendations are tied to specific guideline sources and patient-specific evidence in the record. Every recommendation carries a citation trail; ungrounded recommendations are suppressed.

Workflow

Tool use and escalation

Coordinates non-cognitive actions: notifications, scheduling handoffs, secure messages to clinicians, escalation paths when patient-side events exceed thresholds.

Security & identity

Trust is a continuous plane, not a perimeter.

Identity

Patient-controlled identity with verified linkages

Patient identity is anchored to the patient, with verified linkages to institutional identifiers (MRNs, payer IDs) maintained as references, not as authoritative identity.

Authn / authz

OAuth 2.0 + SMART on FHIR scopes

Every read and every write is scoped, time-bounded, and revocable. Clinician access is scoped by clinical relationship; family delegate access is scoped by the patient.

Consent

Computable consent

Consents are stored as machine-readable artifacts and enforced at the data layer. Research-grade and clinical-grade consents are distinct and individually revocable.

Audit

Append-only audit log

Every read, write, and agent action is recorded with actor, scope, purpose, and outcome. The patient can inspect the log. Tamper-evident by design.

Escalation

Human-in-the-loop boundaries

Defined thresholds trigger clinician notification rather than autonomous action. The agent's authority to act without a human is explicitly bounded and patient-configurable.

Runtime

Cloud-side or on-device, chosen per workflow and per jurisdiction.

Cloud-side

Population-scale reasoning, durable backup
  • Best for cross-institution data assembly
  • Best for guideline updates and shared reasoning improvements
  • Required for population-level research and recruitment
  • Backup-of-record across device loss

On-device

Sovereignty, latency, privacy
  • Best for sensitive reasoning the patient prefers to keep local
  • Best for low-latency conversational interaction
  • Useful for jurisdictions with strict data residency requirements
  • Apple Silicon / mobile NPUs now make frontier-class inference local

Most deployments will be hybrid. The architecture is neutral on this choice: the record and the reasoning layer are designed to run in either mode, and the trust plane enforces the same guarantees in both.

Data provenance

Every fact carries its source.

AI reasoning over patient-controlled records depends on knowing what is authoritative. Guardian Angel tags every datum with provenance and treats provenance as a first-class clinical signal.

Source

Patient-entered, device-captured, EHR-sourced, clinician-validated

The four primary provenance classes. The reasoning layer treats each differently: patient-entered context informs but does not override; clinician-validated facts are load-bearing for downstream recommendations.

Confidence

Measurement quality, not just measurement value

Wearable observations carry sampling rate, device class, and confidence metadata. Manual entries carry timing and free-text context.

Lineage

Derived data points reference their inputs

Agent-derived summaries, alerts, and trajectory findings carry pointers to the underlying observations and the guideline references that grounded them.

Integration

The EHR remains the operational system of record. Guardian Angel sits beside it.

Read

SMART on FHIR app launch and bulk patient access

Where the EHR exposes SMART on FHIR, the Guardian Angel reads on the patient's behalf. Where the EHR exposes bulk patient access under the 21st Century Cures Act, longitudinal ingest is routine.

Write

Structured handoff back to the EHR

Patient-reported outcomes, preference updates, and consented research data flow back to the EHR through standardized structured exchange where the provider system supports it.

Embed

Clinician view embeddable inside the EHR

For clinicians who cannot or will not leave their EHR, the clinician view of the Guardian Angel is itself a SMART on FHIR app launchable in-context. The shared-view property is preserved.

Payer

Claims as a parallel ingest channel

Payer claims data fills gaps in clinical observation and provides cross-institution event detection that EHRs alone cannot offer.