RangerLoop - Raj Singh
Half of imaging referrals never complete.
Core trick
Calls imaging centers in sequence until a slot before the due date is found, then places a guardrailed voice call that verifies identity before booking.
Description
Half of imaging referrals never complete. RangerLoop is a closed-loop care-execution agent that finishes the job. It takes a signed imaging order, or an Abridge ambient encounter, and extracts the study and deadline, every field evidence-linked to the clinician's own words. Then it works the phones: it calls imaging centers one at a time until it finds a slot on or before the due date (striking out slots that are too late), then places a real, guardrailed voice call to the patient to verify identity and confirm - never disclosing clinical results - and books the appointment. It tracks each order ORDERED - SCHEDULED, and escalates safely to a human when it can't proceed.
Claude Opus 4.8 is the visible orchestrator, deciding each next action via tool use with its reasoning streamed live; a thin deterministic policy layer validates every choice (the agent decides WHAT; the policy decides WHETHER), including a safety gate that blocks any booking without verified patient identity. Every record is FHIR-shaped (ServiceRequest / Patient / Task / Appointment), ready to write back to Epic or Oracle Health. Built entirely today; the voice/telephony stack is pre-existing infrastructure consumed over an HTTPS API only.
README excerpt
# 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,