Pressing SOS starts five separate machines: deciding what is wrong, deciding how urgent it is, choosing a hospital, telling that hospital, and moving the people who need to move. This page takes each one apart.
The record reaches all four places from one press · solid lines are pushed automatically, dashed is on demand
Say what is happening, or tap a type. The classifier runs on the phone, so a basement, a lift or a rural road makes no difference. The cloud model is a fallback for genuinely ambiguous phrasing, never a dependency.
The model sits inside the app · the cloud link is dashed because it is optional
no network involved
plus whatever the signal costs
Bars are linear — the on-device bar is not a rounding error, it is 1,400× shorter
Settled instantly, offline, with no errors on the held-out evaluation. The other 39% are the ambiguous phrasings, and those go to the cloud model — which is better at exactly those.
more accurate than ours on raw classification
every case it got wrong, it abstained on instead of guessing
the 61% it chooses to answer
28 phrasings the model had never seen · the point is not beating an 8B model, it is knowing when to stay quiet
A heart attack needs a catheterisation lab. The emergency room four minutes away that will stabilise you and then arrange a transfer has just spent the only minutes that mattered. So distance is the input we weigh least.
Solid line is the route taken · dashed is the nearer hospital that was rejected for having no cath lab
ranked first by distance — then transfers you, and the clock keeps running
ranked first by us: 7 minutes further, and definitive treatment on arrival
closer than the choice, but reported it is not accepting
Illustrative arrangement · the ETAs come from real road-network routing, the capability and diversion flags from what a partner facility reports
Log vitals and the score is arithmetic on published thresholds — the vital-sign subscores of NEWS2, the early-warning score used in hospitals to catch deterioration before it becomes an emergency. Same numbers in, same flag out, every time, with every point traced to the reading that produced it.
Synthetic patient, real thresholds — the bands are NEWS2's published escalation bands, the curve is illustrative
≤8 or ≥25 scores 3
≤91% scores 3
≤90 or ≥220 scores 3
≤40 or ≥131 scores 3
Any single parameter scoring 3 escalates on its own, even with a low total
The unvalidated layer can raise low to medium and never to high — enforced by a test
Stated on every alert
NEWS2 has seven parameters and Relay now captures all seven — but only when you record them. When any are missing the score reports itself as partial and names what is absent. A phone cannot always take your temperature, assess your consciousness level, or know whether you are on supplemental oxygen. A partial score can only ever under-estimate risk, which makes it safe as a reason to seek care and unsafe as a reason not to. Decision support, not a diagnosis.
The moment SOS fires, a shared room opens. Family see live location and the same countdown. A bystander can take over if you stop responding. If a partner hospital is connected, their emergency board has your blood type, allergies and medications before the car moves — and the app only claims “the ER is expecting you” when a real facility actually acknowledged.
The second a panicking family member leaves for another application, the hospital briefing, the arrival estimate and the shared room stop travelling with them. So turn-by-turn is built in — vector tiles, a chase camera, snap-to-route, a wake lock so the screen never dies at a junction, and a night mode.
The full re-scan branch exists because of a real bug — a windowed-only search silently latched onto the wrong part of the line after a GPS jump and under-reported distance by half
A few minutes now is what makes every machine on this page do something useful at 3am.