Watching live. Walk around, then try a controlled drop (a pillow holding the phone works well) to see the rule-based detector fire. Train an AI model in the Train tab to see it predict alongside the rule.
Tip: place the phone down gently on hard surfaces (tables, counters) — a hard drop can still read as a real fall.
Tap a button, then perform the motion for 2 seconds.
How this works
On first launch the app spends 2 seconds calibrating your phone's own resting baseline (tap "recalibrate baseline" any time you change how you're holding it). If you entered a height, it uses a standard waist-height ratio to estimate your phone's typical fall distance and widens or narrows the detection window slightly — a real but modest refinement, since the surface you land on matters far more than fall height and that part we can't measure.
The AI model fuses three sensors — accelerometer, gyroscope rotation, and (if enabled) microphone level — plus whether the phone ends up lying flat, nine features total. It's a small neural network (one hidden layer), not plain logistic regression, written in plain JavaScript and gradient-checked before shipping, since a simpler linear model has a hard ceiling on how well it can separate messy real-world motion no matter how much data you give it. Training combines your real recorded samples (weighted 4x so they dominate) with a simulated bootstrap set covering falls, walking, sitting, hard set-downs, and car braking — not a real imported dataset, since there's no network access here to fetch one and public datasets use different sensor placements anyway, but a stand-in grounded in the same patterns. Recording now waits 1 second after you tap so your reaction time doesn't push the motion to the wrong part of the window, and captures for the same duration the live monitor uses, so training and live detection see the same shape of data.
When a fall is flagged, nothing happens immediately — it first has to survive a 5-second stillness check: the phone must stay within a small tolerance of your baseline (tiny movements like breathing are fine; anything bigger resets it, since that's more consistent with still being upright than being down and not getting up). Only after that does the 15-second "Are you okay?" check-in start.
Every stage now announces itself out loud, not just the check-in: "Fall detected" the moment the 5-second confirmation starts, the full question when check-in begins, and the warning when it escalates. Speaking out loud works reliably everywhere, including iOS Safari. Listening for your spoken answer is best-effort only: browser speech-to-text support is genuinely inconsistent, and iOS Safari specifically has known bugs with it — so the on-screen buttons and every timeout still work no matter what, exactly so voice is a bonus, never the only way to respond.
"Listen for help" (separate toggle, off by default) runs independently of motion entirely — say "help me" or "I'm not okay" at any point and it jumps straight to the check-in, no fall needed first. It listens for full phrases rather than single words like "no" or "help" alone, since without the context of just having asked a question, single common words would misfire on ordinary conversation. It automatically pauses during an active check-in or alert (so it isn't fighting that stage's own listener for the microphone) and resumes right after.
If nothing cancels it, "Notify Emergency Contact" grabs your GPS and opens your phone's share sheet, and (if enabled) the flashlight blinks slowly — a real distress signal, not a rapid strobe, since fast flashing is a genuine photosensitive-seizure risk. Flashlight control only reads the camera's torch capability, never a video frame. Note: it doesn't work in iOS Safari — a real platform limitation, not a bug — but works on Android Chrome.
For car false-positives: record a few "phone braking hard in a cupholder" clips as Normal. A real fall ends in near-total stillness; a car keeps vibrating — the model learns that difference from your data.
Sound is a learned signal now, not a required one. Earlier this hard-required a sound spike, which reliably stopped table set-downs but risked missing real falls that aren't loud — a landing on carpet, a fall with no scream. The AI model still gets peak sound level and how sustained it is as 2 of its 9 features, and is trained on falls across a range of volumes (a quiet carpeted fall, a fall with a cry, a loud tumble), so it can weigh sound when it's a genuine tell without requiring it. The honest tradeoff: the rule-based detector's defense against hard table drops is weaker than it was with the hard gate, since it's back to peak-and-stillness only. Your height still scales the peak bar: a taller person's real fall free-falls further and hits harder, while a table set-down doesn't get harder just because you're taller.