Sentinel
Sentinel — an agentic system that detects clinician burnout before it hits, acts to reduce it, and verifies the load actually dropped.
Core trick
Scores each clinician's EHR activity and schedule load against a personal baseline, then confirms with wearable HRV before recommending one action.
Description
Sentinel — an agentic system that detects clinician burnout before it hits, acts to reduce it, and verifies the load actually dropped.
Sentinel treats the clinician as the patient. Burnout shows up in the data before it becomes reality: rising after-hours charting, growing note backlogs, no recovery days — then suppressed HRV and eroding sleep. Sentinel ingests both sides — EHR activity and schedules as the leading signal, wearable vitals as confirmation — and computes a single Load Index against each clinician's personal baseline. Fully computed, never hardcoded.
It helps rather than informs: one recommended action at a time, quick wins with formula-computed score impact, and unlogged strain spikes detected from wearable telemetry that the clinician can tag in one tap. It acts: drafts the note backlog for sign-off, proposes recovery blocks, and escalates to leadership with consent. A Claude-powered voice copilot reasons over the clinician's full context — ask why the score moved, negotiate the schedule, send a coverage request.
Privacy is the architecture: wearable and personal health data shape the score but never leave the clinician's view — the hospital sees only the composite score and workload signals it already owns. And Sentinel closes the loop: it re-checks the data days later and verifies the load actually dropped.
README excerpt
# Sentinel **The clinician is the patient.** Sentinel is an agentic system that watches the data a hospital already generates (EHR activity, scheduling) plus the clinician's own private signals (wearable, personal health record), detects burnout **before it lands** — the record leads, the body confirms — then **acts**: drafts the note backlog, triages the inbox, renegotiates the schedule by voice, escalates with consent behind a hard privacy wall, and verifies the load actually dropped. Built in one day at the Abridge × Anthropic × Lightspeed hackathon (July 18, 2026). **All data is synthetic**, seeded to published occupational-stress effect sizes; integrations are mocked interfaces mirroring real products (Epic Signal, QGenda, Apple HealthKit, pharmacy fill feeds). ## Run it ```bash npm install npm run dev # → http://localhost:3000 ``` Works fully offline out of the box (curated agent outputs). To run the agents live on Claude: ```bash cp .env.example .env.local # add your ANTHROPIC_API_KEY ``` `DEMO_MODE=cached` forces offline mode even with a key. ## The demo - `/` — the clinician's view: Load Index, top drivers, "Tell me more," talk to Sentinel (voice) - `/insights` — full trajectory, drivers by privacy tier, live agent feed (Reasoner vs. Skeptic), sign-off actions, the Care Plan - `/admin` — the CMO's view: roster risk, escalation inbox — and everything private locked away - Press **`d`** to toggle the demo timeline (scrub Jun 28 → Jul 25, jump to Trigg