RangerLoop

A closed-loop care-execution agent: a clinician's signed imaging order (or an ambient clinical encounter) → extract the action + due date (evidence-linked) → call imaging centers for a compliant slot → call the patient to confirm & book (guardrailed, never discloses results) → track ORDERED → SCHEDULED → escalate safely on failure.

Ambient AI captures the visit. RangerLoop gets it done.

Live demo: https://rangerloop-api.fly.dev — the board shows example loops across states. Click "＋ Imaging order" or "Abridge encounter" to run a fresh loop end-to-end (imaging-center calls are simulated; the patient leg places a real, guardrailed voice call). Demo video: https://www.loom.com/share/0d5d94ef03ba45988e6a5d908fa61913

Built at the Abridge × Anthropic × Lightspeed healthcare hackathon (2026-07-18).

Built today (in this repo)

• Evidence-linked extraction from a signed order or an Abridge ambient-FHIR encounter — every action-driving field (study, due date, phone) is tied to an exact quote; for encounters, the quote is the clinician's spoken transcript line.
• The loop state machine — a forward-only reducer that is the single writer of all state, timeline, slots, and escalations.
• The Claude tool-use orchestrator — decides each next action; its reasoning renders live on the dashboard.
• The deterministic policy / safety layer (calls are sequential, no patient call before a compliant slot, no re-dialing a found center, mandatory escalation) and healthcare guardrails (never disclose results / findings / diagnosis; verify patient identity by date of birth).
• The real-time dashboard: loop board, imaging-center campaign view, per-call conversation, Claude reasoning cards, evidence, FHIR resources, escalation queue, and an operator Stop.

Pre-existing infrastructure (disclosed, not in this repo)

The AskRanger voice/telephony stack (placing/answering calls, IVR navigation, the live voice agent, structured field capture, event webhooks) is pre-existing, consumed over an HTTPS API only. None of its code is in this repository.

Architecture

• Claude is the visible orchestrator (tool-use): it sees the loop state + tools [call_imaging_center, call_patient, send_sms_confirmation, escalate_to_human, mark_scheduled] and decides each next action; its reasoning is streamed to the dashboard.
• A thin deterministic policy underneath validates every choice. The agent decides WHAT, the policy decides WHETHER.
• FHIR-shaped records (ServiceRequest / Patient / Task / Appointment) — the loop's state is native EHR data, ready to write back to Epic / Oracle Health.
• Real-time React dashboard polling the backend every 2s.
• Stack: TypeScript · Node + Fastify · React + Vite · Claude Opus 4.8 (adaptive thinking + tool use).

Run it locally

Requires Node 22+ and an ANTHROPIC_API_KEY (extraction and the orchestrator use Claude).

npm install
npm run build
ANTHROPIC_API_KEY=sk-... ASKRANGER_MODE=mock node server/dist/index.js

Open http://localhost:8080 and click "＋ Imaging order". ASKRANGER_MODE=mock simulates the calls (phone-free, deterministic); set ASKRANGER_MODE=live with the AskRanger API credentials to place real calls through the disclosed voice stack.

Notes

• The SMS confirmation step is stubbed (the SMS endpoint is a deferred integration). Every other step — extraction, the imaging-center campaign, the guardrailed patient voice call, and the FHIR booking — runs for real.