Under the hood
Under the hood

How it tells a fall from a dropped book

Two radars, two microphones and our fall model check every possible fall several times before your phone rings. Here is the hardware, how it decides, what leaves her home, and how we test it.

No camera, no lensNothing to wear or chargeNo recording on a normal day

Written by the team that builds the Nightlight · Prototype specs as of Sept. 29, 2026 · Drawings and animations are illustrations, not sensor recordings

Side section of the Nightlight prototype on a wall: two microphones at the top, a 60 GHz radar tilted down, a warm light band, a 24 GHz radar, the processor at the back, a speaker that fires down, and a wall plate with adhesive strips. 72 mm 01 Two microphones the thud, then her answer 02 60 GHz radar tilted down at the floor 03 Warm light band 2200–2700 K, no flicker 04 Processor runs the fall model 05 24 GHz radar where people are 06 Speaker fires down, tuned for speech 07 Wall plate 2 adhesive strips 1234567
Side section · prototype · 150 × 88 × 72 mm
60 + 24 GHz

Two radars. One sees the fall, the other knows where people are.

0 cameras

No lens anywhere. Nothing on it can take a picture.

4 checks

Before your phone rings, from the first reading to a second look.

60 seconds

From a fall to your phone, when she can't answer.

01

From a radar echo to your phone

Every alert passes through four layers. The first two live on the Nightlight, in her home. Only three things ever cross the line out, and each one only at a certain moment.

Her home
Layer 1 · SenseIn the room
60 GHz radar · the fall24 GHz radar · where people are2 microphones · the thud, then her answer

Three kinds of signal, read many times a second. None of them can make a picture.

Layer 2 · DecideOn the Nightlight
Quad-core processorOur fall modelSpeaker and light

Our model weighs the three signals over the same few seconds. If they look like a fall, the Nightlight asks her out loud, and the radars take a second look. No single signal can call anyone.

What crosses the line, encrypted
  • An alertThe room, the time, a simple radar shape and which checks it passed.
  • Her spoken answerA few seconds of sound after a possible fall, sent to be understood.
  • A live callOnly during an alert, and only if you start it.
Layer 3 · ReachIn the cloud
Her answer, understoodHer circle, in orderA schedule that survives restarts

Her answer becomes one of three results: she's fine, she needs help, or no clear answer. If she needs help or can't answer, it works through her circle one person at a time until someone picks up.

Layer 4 · YouYour phone
An alert in the appA regular phone call and a textTalk to her through the Nightlight911 with one tap

You see the room, the time and the radar shape. Whoever answers decides what happens next.

02

What's inside, in numbers

These are the specs of the prototype on our bench today. Numbers marked target are what we design and test against. We'll update this page as they change.

Sensor 1 · 60 GHz radar

Sees the fall

60 GHz FMCW120° × 100° viewLooks down at the floor

It sends out radio waves and times the echoes. That gives it distance and speed: enough to see a body drop fast and stay low, and to notice the small movements a pillow never makes. Not enough to make a picture.

Processed: on the Nightlight. Nothing it reads leaves her home.
Sensor 2 · 24 GHz radar

Knows where people are

Up to 3 peopleUp to 6 m away120° × 70° view

It follows where each person is and how fast they move. So it can tell a fall from something falling near her, and it knows whether she is alone in the room.

Processed: on the Nightlight. Nothing it reads leaves her home.
Sensor 3 · Two microphones

Hear the thud, then her answer

2 mics, set apartLoudness and shape onlyAnswer: a few seconds

Day to day they only measure how loud a sound is and its shape. Nothing is recorded. After a possible fall, the Nightlight records her answer for a few seconds. Two microphones, set apart, also tell which way a voice comes from.

Processed: everyday sound on the Nightlight. Her answer is understood in the cloud and not kept.
The brain · Processor

Runs the fall model

Quad-core, 64-bitWi-Fi 5Bluetooth 5.0

A small computer inside the Nightlight runs our fall model and every check before a call. It reaches your phone over Wi-Fi. Bluetooth is only for setup.

Processed: the fall checks run here, in her home, not on our servers.

What a radar actually gets

A radar does not see a room the way a camera does. It gets reflections: points with a distance and a speed. Researchers who reviewed 74 papers on radar fall detection describe the data as patterns of distance, speed and direction of movement.1 It cannot show what she is wearing, what she is reading, or her face.

Someone walks in. The sensor gets points that move.

An illustration, not a recording. Each dot stands for a reflection with a distance and a speed.

03

How the AI decides

Each sensor alone gets things wrong. A radar can take a cat for a person. A thud can be a dropped pan. So the model acts only when the signals agree, at the same moment, in the same spot. Pick a moment and watch it play out.

60 GHz radarbody height
standingfloor
24 GHz radara person at that spot
Microphonesloudness
MovementWaiting
A person thereWaiting
Her answerWaiting
Second lookWaiting

    An illustration of the logic, not sensor recordings. All five moments are in our test set.

    MovementDrops that aren't a body falling: a book, a pan, a towel.Nothing happens.
    A person thereNobody is where the fall seemed to happen.Nothing happens.
    Her answerShe's fine. She sat down hard or dropped something.It ends at home.
    Second lookShe got up on her own.It ends at home.
    04

    Before anyone is called

    The first minute belongs to the Nightlight. After that, it wakes people up, family first.

    1. 0:00

      A possible fall

      Movement, presence and sound point the same way.

    2. 0:05

      It asks her

      “Mom, are you okay?” If she says she's fine, it ends here and nobody is told.

    3. 0:35

      Still no answer

      It has asked twice. Now it gets ready to call.

    4. 1:00

      Your phone rings

      After a second look, and only if the radar still sees her on the floor. You get the room and the time.

    5. 5:00

      The next person

      No answer from you? It calls the next person in her circle, with a text as backup.

    6. Then

      A person decides

      Go over, call a neighbor, or call 911 with one tap.

    05

    What stays in her home, and what leaves

    Here is every piece of data the Nightlight sends, and when.

    Stays on the Nightlight

    Everything it senses

    • What the radars readPositions, speeds, the fall itself.
    • Everyday soundOnly loudness and shape. Nothing is recorded.
    • The fall checksThey run on the Nightlight's own processor.
    Leaves it, encrypted

    Only these, and only then

    • All the timeA small “still online” check-inSo your app can show it's working. No sensor data, no sound.
    • After a possible fallHer spoken answerA few seconds, sent to be understood. Used only to work out her answer. We don't keep it. It works out what she said, never who is speaking.
    • In an alertThe room, the time and a radar shapePlus which checks it passed. Never a picture.
    • If you start itA live call with herThrough the Nightlight, during an alert.

    What you get: the room, the time and a shape from the radar. Never a picture.

    NeverA camera or videoVoiceprintsHer data used for adsHer data sold
    06

    How we test it

    We test against a fixed set of moments on our bench. Falls are staged on a mat, with a spotter. Each one is repeated many times, from more than one spot in the room.

    9kinds of falls

    1. A hard fall forward
    2. A fall to the side
    3. A fall backward
    4. Rolling out of bed
    5. A slow slide down the wall
    6. Slumping from a chair to the floor
    7. A fall partly hidden by furniture
    8. A fall, then lying still for 90 seconds or more
    9. A fall, then getting up after about 20 seconds

    11things that look like one

    1. A blanket or pillow dropped on the floor
    2. A cat or a dog
    3. Sitting or lying on the floor on purpose
    4. Bending down to tie a shoe
    5. Sitting still in a chair for 10 minutes
    6. A fan and a moving curtain
    7. A robot vacuum going by
    8. A loud TV or radio
    9. Two people, one lies down on the bed
    10. An empty room, all night
    11. A normal night in bed, up to the bathroom and back
    Falls caughtBy kind of fall, by distance and by angle.
    False alertsEvery time it would have woken you for nothing.
    Seconds to your phoneFrom the fall to the first call.

    We run the set twice: with the Nightlight where we tell you to mount it, and in the worst spot a real home might force.

    Before betaZero missed long lies on the bench

    A long lie, on the floor and unable to get up, is the case this product exists for.

    For launchFewer than one false alert per home per month

    A target, not a result yet. We'll post the bench results on this page, misses included.

    Why real homes are the hard part
    2 a week, 8.6 a day

    False alarms in two studies of radar in older adults' homes, from a review of 74 papers. Both caught every fall they saw.1

    98.5%

    of 130 fall detection studies tested only staged falls.2 A staged fall by a younger adult is not a real fall at 85. That is why our beta runs in real homes.

    07

    Where the build is

    The Nightlight is not on sale yet. Here is what already works and what comes next.

    1. Jul 2026 · done
      Working prototype

      The fall radar, the spoken check and the alert in the family app work end to end.

    2. Now
      All three signals

      Adding the second radar and the sound check to the fall model.

    3. Now
      Understanding her answer

      Fine, help or silence, in her own words.

    Updated Sept. 29, 2026

    08

    Specs

    Prototype, September 2026. Values marked target are design goals we test against. Specs may change before shipping.

    Sensing

    Fall radar
    60 GHz FMCW, 120° × 100° view, tilted down toward the floor
    Presence radar
    24 GHz FMCW, up to 3 people, up to 6 m, 120° × 70° view
    Microphones
    2, set apart to tell which way a voice comes from
    Camera
    None. No lens.

    Voice

    Speaker
    Full-range, faces down through a slot, tuned for clear speech
    Loudness
    80 dB at 1 m target
    Spoken check
    “Are you okay?” in a calm voice. Her answer is understood in the cloud.

    Light

    Color
    2200–2700 K, warm white
    Dimming
    Flicker-free
    Night mode
    Turns on when she gets up at night

    Processing

    Processor
    Quad-core, 64-bit
    Fall checks
    On the device

    Connection

    Wi-Fi
    Wi-Fi 5
    Bluetooth
    5.0, for setup only
    Setup
    About 5 minutes, from your phone. Nothing for her to set up.

    Power

    Input
    USB-C, 5 V, from a wall outlet. Nothing to charge.
    Battery
    None. Without power it can't alert anyone.

    Physical

    Size
    150 × 88 × 72 mm
    Weight
    300–400 g target
    Shell
    Matte ceramic, opal light band
    Mounting
    On the wall, about 180 cm up. Two adhesive strips, no drill.

    Privacy and safety

    Her consent
    Fall detection turns on only after she says yes
    In transit
    Encrypted
    Voice
    Understands what she says, never who is speaking
    Certification
    FCC, before it ships
    What it can't do
    • It can't prevent a fall, and it can't catch every fall.
    • It covers the room it's in. For the bedroom and the bathroom, that means two Nightlights.
    • It can miss a fall hidden behind furniture. Where you mount it matters.
    • It follows up to 3 people in a room. A busy room is harder.
    • Without power or Wi-Fi, it can't reach your phone.
    • It is not a medical device and not a substitute for 911.

    Sources

    1. Hu S, Cao S, Toosizadeh N, Barton J, Hector MG, Fain MJ. A survey on radar-based fall detection. IEEE Robotics & Automation Magazine, 2024 (preprint: arXiv 2312.04037). Table II: Liu et al. and Su et al., real-life senior resident activities. arxiv.org/abs/2312.04037↑ Back to the text
    2. Ishaq M, Guastella DC, Sutera G, Muscato G. A systematic review of fall detection and prediction technologies for older adults. Applied Sciences 16(4):1929, 2026. doi.org/10.3390/app16041929↑ Back to the text

    If someone is hurt or cannot get up, call 911.