Skip to content

CareLoop AI

Healthcare Commitment Intelligence & Closed-Loop Care — from a doctor’s word to a verified outcome.

This site publishes the artefacts behind the CareLoop AI POC review. They are all generated from the same repository revision, so the diagram, the numbers and the deck describe one system.

Start here

  • Interactive architecture

    The whole system as one clickable diagram. Every node carries the file and line that proves it, and it renders in both light and dark mode.

  • Architecture notes

    The same system in words: four subsystems, where each one ends, and the honest limits of the POC.

  • Animated slideshow

    The six-slide review deck. Arrow keys advance, P opens presenter mode with speaker notes, ESC shows the slide index.

  • Static plate (2× PNG)

    The same diagram as a print-ready raster, for slides and documents.

What the demo does

A doctor pastes a note. One agent writes typed care cards — medication, test, referral, next visit, general task — with due dates read straight out of the note. Cards carry dependencies, so a card whose blocker is still open turns at_risk on its own and heals the moment the blocker closes. A 28-diagnosis checklist asks the model which expected items this note’s cards miss, and offers them to the doctor to approve or dismiss. An MCP server exposes the same records to agents over six careloop_* tools, with no database credentials.

Verified on the demo database

Measure Value
Notes processed end to end 24
Cards saved 63
Care-gap recommendations acted on 21
Cards closed with evidence 9
Patient messages triaged 98 (26 urgent, 72 not urgent)
Urgent-item worst-case wait 10 s (watcher poll interval)
MCP client test suite 5 passed

Counted directly from the demo Postgres volume on 7 October 2026. These are rows you can query, not a benchmark.

Honest limits

No EHR or FHIR feed — notes are pasted by hand. One shared demo password, no per-user tokens on the MCP server. Urgency triage results are written to their own schema but are not surfaced in the UI yet. There is no formal eval harness; the five passing tests cover the MCP client, not the model.