Stanford Med · Unit 5B Medicine·Day shift·Mon · 9:42 AM

System architecture

One platform · ingestion + core + outbound + UI · explicit build vs. partner
Architectural thesis

OpenBed is the coordination layer — not the EHR, not the SNF directory, not the prior-auth submitter.

We sit on top of Epic / Cerner via FHIR, automate the drafting into CoverMyMeds / Availity, surface SNF capacity from Naviguide / WellSky, orchestrate the human team, and audit every action. The specialized vendors are best-in-class at their slice; we are best-in-class at making them work as one operating layer.

14 build · core IP18 partner · commodity infra1 build on (Anthropic, Inngest)13 live · 15 pilot · 5 planned
OpenBed AI · platform architecture

One coordination layer · four tiers · 30+ components

Ingestion · read from source systems
Mostly partner · 11 components
Epic FHIR R4
Epic Systems
last sync · 12s ago
Redox (multi-EHR connector)
Redox
last sync ·
HL7 v2 ADT listener
HL7 v2.x (A01, A02, A03, A08)
last sync · Stream
Eligibility (X12 270/271)
Availity (primary), Change Healthcare (backup)
last sync · 2 min ago
Surescripts (formulary + benefit)
Surescripts
last sync · 8 min ago
Naviguide capacity feed
Naviguide (or WellSky CarePort / Aidin per hospital)
last sync · 5 min ago
DME partner capacity (Lincare, Apria, Inogen)
Lincare, Apria, Inogen, Parachute Health
last sync ·
ModivCare NEMT availability
ModivCare
last sync · 1 min ago
Patient portal events
Internal SSE + REST
last sync · Stream
Voalte ack stream
Vocera/Voalte/Halo
Brain tick (scheduled detectors)
Inngest cron
last sync · 30s ago
OpenBed Core · the coordination IP
We build · 8 components
Six-Dimension Readiness Engine
Internal
PaCHE barrier ontology + FSM
Internal
Brain detectors
Internal
AI agent layer (Claude via runAgent)
Anthropic
Resource matcher (graph)
Internal
Task orchestrator
Internal
Audit spine (AgentInference + Submission)
Internal
Event bus (Inngest workflows)
Inngest
Outbound action · write to external systems
Mostly partner · 9 components
EHR writeback
Epic Systems
CoverMyMeds (PBM prior auth)
CoverMyMeds (Surescripts)
Availity (medical PA + EDI)
Availity
Naviguide e-referral
Naviguide
Parachute Health (DME orders)
Parachute Health
ModivCare NEMT booking
ModivCare
Twilio + Vapi/Retell (voice agents)
Twilio + Vapi
Documo (e-fax fallback)
Documo (or Phaxio)
Notification bus
Twilio + SendGrid
UI surfaces · where humans interact
We build (mostly) · 5 components
CM Console
Internal
MD Widget (Epic Sidebar)
SMART-on-FHIR app
Patient + Family Portal
Internal
Voalte-style task feed
Vocera/Voalte/Halo
Admin / Throughput dashboard
Internal
Build (core IP)Partner / integrateBuild on platformRing color = status (green live · amber pilot)
Build vs. partner matrix

Explicit posture for every component

Build (core IP)

These are the components no vendor has. They are the moat.

  • HL7 v2 ADT listener
    Pilot integration
    Legacy ADT is the most reliable real-time signal. We listen via Mirth Connect-style integration engine inside our infra.
  • Patient portal events
    Live in prototype
    We own the portal end-to-end; events stream to the engine in <500ms via SSE.
  • Brain tick (scheduled detectors)
    Live in prototype
    Domain-specific signal engine. Scheduled re-evaluation runs deterministic detectors; AI generates narrative on detail view only.
  • Six-Dimension Readiness Engine
    Live in prototype
    Core IP. No vendor has this. Recomputes on every relevant event; emits pillar deltas to the routing engine.
  • PaCHE barrier ontology + FSM
    Live in prototype
    The substrate of the platform. Every barrier classified, owned, state-tracked, audited.
  • Brain detectors
    Live in prototype
    Severity is computed deterministically. AI generates narrative only on demand via runAgent (never decides severity).
  • Resource matcher (graph)
    Live in prototype
    Deterministic ranking on top of curated Resource Hub. Where AI assists, it only generates rationale text.
  • Task orchestrator
    Live in prototype
    Cross-role coordination is the moat. No off-the-shelf workflow tool understands hospital roles + state semantics.
  • Audit spine (AgentInference + Submission)
    Live in prototype
    Legally-load-bearing. Answers payer dispute, regulator audit, plaintiff lawyer — 'what did the AI propose, who approved, what was sent.'
  • Notification bus
    Live in prototype
    Custom routing logic per role / urgency / channel. Wraps Twilio for SMS, SendGrid for email.
  • CM Console
    Live in prototype
    Hero surface. Built on Next.js + Tailwind; deep integration with the engine.
  • MD Widget (Epic Sidebar)
    Pilot integration
    SMART-on-FHIR launched inside Epic. Surfaces the engine where MDs work.
  • Patient + Family Portal
    Live in prototype
    No vendor owns this surface for discharge specifically.
  • Admin / Throughput dashboard
    Live in prototype
    Required for CFO + hospital ops buyer. Built on Recharts; reads from audit spine + engine state.
Partner / integrate

Commodity infrastructure. We integrate, we don't rebuild.

  • Epic FHIR R4
    Pilot integration
    EHR connectivity is commodity infrastructure. Building our own connector is an 18-month rabbit hole. We integrate via Epic's FHIR R4 API + SMART-on-FHIR launch.
  • Redox (multi-EHR connector)
    Planned
    Multi-EHR support without per-vendor engineering. Redox handles auth, mapping, and rate-limits.
  • Eligibility (X12 270/271)
    Pilot integration
    EDI clearinghouse is commodity. Availity is the largest; Change Healthcare and Waystar are alternatives.
  • Surescripts (formulary + benefit)
    Pilot integration
    De-facto standard for ePrescribing infrastructure. We use the read side for formulary + the write side for ePrescriptions.
  • Naviguide capacity feed
    Pilot integration
    Post-acute capacity is the hardest signal to get without a vendor. Naviguide / WellSky CarePort / Aidin all aggregate it; pick by hospital preference.
  • DME partner capacity (Lincare, Apria, Inogen)
    Planned
    Direct vendor APIs where available; Parachute Health for aggregation. Voice agent fallback for vendors without APIs.
  • ModivCare NEMT availability
    Pilot integration
    Largest CA Medicaid NEMT broker. Direct API integration.
  • Voalte ack stream
    Planned
    Most hospitals already on Vocera/Voalte/Halo for nursing comms. We listen to their event stream rather than asking them to switch.
  • Event bus (Inngest workflows)
    Pilot integration
    Durable execution + retries + observability without building our own. Inngest is the right primitive for this shape of work.
  • EHR writeback
    Pilot integration
    Closes the loop. Every CM action that should live in the EHR is written via FHIR R4.
  • CoverMyMeds (PBM prior auth)
    Pilot integration
    CoverMyMeds is the de-facto standard with PBM connections (CVS Caremark, Express Scripts, OptumRx). We automate the drafting; they own the pipe to the PBM.
  • Availity (medical PA + EDI)
    Pilot integration
    Largest provider portal + payer connectivity in US.
  • Naviguide e-referral
    Pilot integration
    We don't replace the SNF e-referral vendor — we automate the packet drafting and surface response in the CM workflow.
  • Parachute Health (DME orders)
    Planned
    Parachute aggregates DME vendors; we send one order, they fan out.
  • ModivCare NEMT booking
    Pilot integration
    Direct broker booking.
  • Twilio + Vapi/Retell (voice agents)
    Pilot integration
    Twilio for telephony infra; Vapi/Retell for the voice-AI runtime. We compose prompts via runAgent.
  • Documo (e-fax fallback)
    Pilot integration
    Unfortunately required in healthcare. Documo is HIPAA-safe.
  • Voalte-style task feed
    Planned
    We push tasks into Vocera/Voalte rather than asking nurses to switch apps.
Build on platform

Foundational platforms we leverage (LLM runtime, durable workflows).

  • AI agent layer (Claude via runAgent)
    Live in prototype
    Built on Anthropic Claude (Sonnet 4.6 + Haiku 4.5). Every call routed through runAgent — writes AgentInference row with cost, latency, prompt hash, approval triple.
How this screen works
This page is the strategy map · static + linked from every other screen
Inputs
What this screen reads
  • src/lib/architecture/components.ts (typed component registry)
  • Build vs. partner classification per component
  • Live status per component (synthetic in prototype)
Engine
What it computes
  • Renders 4-layer diagram with posture + status indicators
  • Groups components by posture for the matrix view
  • No live computation — strategic document
Outputs
What it writes / routes
  • Cross-links to /cm/integrations for partner detail
  • Cross-links to /cm/dynamic-engine for runtime view
Refresh trigger
When it updates
  • Edits to src/lib/architecture/components.ts
  • Manual review on each major architectural decision
Partners involved:Strategy doc · references all partners listed in /cm/integrations
Strategic note

Why not build the whole stack?

Healthcare integration is commodity. Spending 18 months building an Epic FHIR connector loses the discharge-coordination race to whoever ships a platform layer faster.

Strategic note

Why not just install Naviguide + CoverMyMeds?

They don't talk to each other. None of them surface the six readiness dimensions. None of them have the patient portal. None coordinate the human team. That's the moat.

Strategic note

What ships in pilot vs. design partner #2?

Pilot: Epic FHIR + CoverMyMeds + Naviguide + Lincare/Apria + ModivCare + Twilio. Pilot #2: Cerner via Redox + Parachute Health + Vocera/Voalte writeback + Inngest at-scale workflows.

Demo data · no PHI · mocked Epic + payer endpoints