Five things start at once. The phone works out what is wrong and how urgent it is, picks a hospital, sends that hospital your details, and gets your family and a driver moving. Each section below covers one of them.
The hospital step only happens if the hospital is a Relay partner. Everything else happens for everyone.
You can say what is happening or tap a type. The classifier runs on the phone itself, so it still works in a basement, a lift or out in the country. If the wording is ambiguous it asks a larger cloud model, and if there is no signal for that, it goes with its own best guess.
Most cases finish on the phone. The dashed path is only used when the phone is not confident.
no network involved
plus whatever the signal costs
Both bars are on the same linear scale. The on-device one really is about 1,400 times shorter.
61% of test cases were settled on the phone, offline, with no errors on the held-out set. The other 39% were the ambiguous ones, and the cloud model handles those better, so they go there.
more accurate than ours on raw classification
when it would have been wrong, it handed off instead of guessing
the 61% it chooses to answer
Tested on 28 phrasings the model had not seen before. The 8B cloud model is more accurate overall. Ours is built to hand off when it is unsure.
A heart attack needs a catheterisation lab. If the nearest ER does not have one, it will stabilise you and then transfer you, and that transfer uses up the minutes that count. So distance gets the least weight in the ranking.
An illustrative example. Drive times come from road-network routing. Cath lab and diversion status come from what each hospital reports.
When you log vitals, the score is plain arithmetic on published thresholds: the vital-sign parts of NEWS2, the early-warning score hospitals use to spot patients who are getting worse. The same numbers always give the same result, and every point can be traced back to the reading behind it.
Real NEWS2 thresholds, made-up readings. Readings 3, 4 and 5 each count as low on their own, but put side by side they are already climbing two readings before the score reaches medium.
≤8 or ≥25 scores 3
≤91% scores 3
≤90 or ≥220 scores 3
≤40 or ≥131 scores 3
Any one of these scoring 3 escalates on its own, even if the total is low.
The personal baseline is not clinically validated, so it can raise a low score to medium but never to high. A test enforces this.
Shown on every alert
NEWS2 uses seven measurements. Relay can record all seven, but only the ones you enter, and a phone cannot take your temperature, judge your level of consciousness, or tell whether you are on oxygen. When anything is missing, the score is marked partial and lists what is missing. A partial score can only under-estimate risk. If it says high, get help. If it says low, that is not a reason to skip getting help. Relay supports a decision; it does not diagnose.
When SOS goes off, a shared room opens. Family see your live location and the same countdown you do. If you stop responding, someone with you can take over. If the hospital is a connected partner, its emergency board gets your blood type, allergies and medications before you set off. The app only says the ER is expecting you once a real person there has confirmed it.
If a frightened family member jumps to a separate maps app, the hospital brief, the arrival time and the shared room get left behind. So turn-by-turn lives inside Relay: vector maps, a camera that follows the car, snap-to-route, a wake lock so the screen stays on at junctions, and a night mode.
The re-scan branch is there because of a real bug: after a GPS jump, the old search locked onto the wrong part of the route and reported about half the real distance.
It takes a few minutes. Your profile is what the hospital, your family and the triage model all work from, so it has to exist before the night you need it.