Camily — the fall-detection layer for phones already in your home

How it works

Simple gestures. Smart analysis. Real-time alerts.

1
Room phone monitoring

A spare phone stays plugged in and watches one room. The “camera is on” strip shows exactly what it sees.
2
Portal tiles

Every room becomes a tile that refreshes about every 15 seconds, live heartbeats included.
3
Guardian app

Home tiles and history rows are all live, so family and staff see the same truth at the same time.
4
Live tests

Wave and “person in view” flips on. Raise a hand and a HELP incident fires. Drop fast and stay down and a possible-fall check starts.
5
Alert flow

Device detects (≤2s tick) → outbox posts immediately → portal shows it (~15s poll) → guardian gets the notification.

Detection

Can it really see a fallen person?

Yes, as a heuristic. It detects rapid downward movement followed by persistent low posture, then posts a possible-fall incident through the durable outbox.

Rapid drophips fall ≥12% of frame height inside a 2s window
Low posture persistsat least 3 frames inside a 6s window
Incident postedstraight to the outbox, not waiting on the heartbeat

It is deliberately a “worth checking” signal, not a certified fall alarm. Camily refuses to overclaim.

An alert screen with call contact, who is handling it, and a step-by-step timeline — Every alert shows who is handling it and what happened, step by step.

History view listing checks, alerts and who acknowledged them — 62 quiet checks, the alerts that mattered, and who saw them.

Speed

How long to raise the alarm?

Pose inference runs every ~500ms (~2fps), 250ms bursts while a candidate is active
Confirmation needs the person down for ≥3 frames inside a 6s window, plus a short dwell
The incident posts immediately through the outbox, not on the 60s heartbeat
The portal shows it on the ~15s poll; the guardian gets it via the notification worker

Realistic total: under ~30 seconds from hitting the floor to visible on the portal. A few seconds device-side.

It requires vision and zone calibration. If it cannot see, it closes silently as unobservable rather than guessing.

Privacy

How often does it take pictures? Never.

Camera frames are analysed in RAM. In the strict-privacy build the JPEG encoders are physically removed from the binary, and the incident payload literally says “No routine image was uploaded.”

No video feed exists to leak, record or sell
Thumbnails on tiles are bundled synthetic assets, not captures
“Request current view” is closed under strict mode
Words-only alerts: what was seen, not a picture of it

What Safely sees screen showing a sample frame illustration and a No saved camera images badge — “What Camily sees” is an illustration, because there is no photo to show.

Accuracy

Floor exercise or fallen? The differentiators.

Speed of descent

A fall drops the hips fast. Sitting or lying down on purpose is slower and reads differently.

Persistence

The person must stay down across ≥3 frames in 6s. A brief sit-down will not count.

Recovery

Getting back up marks recovery observed and attention is suppressed.

Zones

Calibrated room zones: the bed is expected, the floor zone is not.

Occupancy

An upright person still in frame changes the assessment.

Honest limits

What it cannot do

A fast burpee can mimic the fall signature
A slow slump off a chair (a common real fall) can miss the “rapid drop” trigger and land as a weaker lying expected signal
Dark, occluded or out-of-view means unobservable: it says so instead of guessing confidently

That is exactly why the HELP button and the raised-hand gesture exist: the person can always call out, and Camily never pretends the passive check saw more than it did. Camily is not a medical device and never calls emergency services itself.

Alerts and escalation settings showing quiet hours, blur options and the honest escalation note — Escalation settings say it plainly: automatic calls are not enabled; a person decides what happens next.

For care teams

A provider portal that respects the same rules.

Organisations see the same honest states families do: who needs a response, which rooms are offline, and what is simply simulated or unverifiable. Privacy filtering applies to staff too.

Provider dashboard: eight residents, needs-response queue, site filters and a simulation badge — One glance: residents, open alerts, offline rooms.

Participant directory with a synthetic sample room observation — Participant directory: preferences, contacts, room context.

Operations overview across three homes with coverage and coordinator queue — Operations across homes: coverage, leave, on-call.

Explore the live care-workspace preview →

Screenshots

Real screens, honest states. No mockups.

Everything below is the actual app and portal running against the hosted demo. All names and residents are synthetic.

Home: three rooms, an open alert, last check time — Home: every room, every state, at a glance.

What the room phone sees: illustration only, no saved images — No saved camera images, by design.

Alert detail with call button and timeline — Alerts name who is handling them.

Consent wizard: who is being watched and what they agreed to — Consent is a wizard, not a checkbox.

Household members with guardian and owner roles — Households with clear roles and access.

Who to call: after-hours nurse and family contact — Who an alert asks you to call.

History list of checks and alerts — A history you can actually audit.

Image retention settings: 24 hours, 7 days, 30 days — Retention you control.

Choosing the phone's role: guardian or room phone — Any spare phone becomes the room phone.

Try it on the phones you already own.

One phone stays in the room. Yours gets the alerts. The demo data is entirely synthetic, so you can break things freely.

Join on TestFlight Request the Android build Open the portal demo

Camily is a monitoring and notification aid. It reports what it can see, not a diagnosis. It does not replace a person checking, and it never calls emergency services by itself.