The drone
LOOK-AHEAD
- Flies about 30 cm above and behind the walker's head.
- Sees what a cane can't: the crowd at the corner, the crossing, the car turning in.
Without it: you only learn about the street one cane-length at a time.
a guide dog that flies
Hear the street
before you reach it.
The cane finds the pole when you hit it. Kite tells you first.
A small drone flies just above and behind a blind walker's head, where it can see the route, the crowds and the crossings ahead. A camera on their chest watches the ground at their feet. Kite turns all of it into sound in their AirPods: beeps that come from where things are, and a calm voice that stops them before trouble, then tells them when to go.
- WHAT'S AHEAD
- found by the cane, one step at a time
- seen from above, half a block out
- WHAT'S CLOSE
- the pole you walk into
- a beep that speeds up first
- CROSSINGS
- guessed from traffic noise
- "Stop." then "Walk signal is on."
put this on in headphones, it's 41 seconds
RECORDED FROM THE REAL EARPIECE · LYON STREET, SF (SIMULATED)
One walk to the Palace of Fine Arts
Press play. Every sound you hear came out of Kite's earpiece code, with the real Cartesia voice.
Transcript of what Kite says
Every sound means one thing.
You learn Kite in a minute, the way you learn a turn signal. Where a sound comes from is where the thing is. What it sounds like is what it is. How fast it repeats is how soon you'd reach it.
WHERE
Direction
Rendered with a model of the human head, so each ear hears it a fraction of a millisecond apart.
WHAT
A sound per thing
A soft knock, a rough buzz, a clean pip, a falling glide. Never more than two at once.
HOW SOON
Getting closer
Beeps start about three and a half seconds before you'd meet, then speed up and rise in pitch. At the last second Kite says "Stop."
WHICH WAY
The beacon
A tick from the direction of your next turn. Bright when you're facing it.
Three ways of seeing.
One voice.
A guide dog does a few jobs very well: it stops at curbs and crossings, leads around trouble, watches overhead, and refuses to go when a car is coming. Each part of Kite takes one of those jobs.
The chest camera
REFLEXES
- Watches the ground at the walker's feet, which the drone can't see.
- Catches the sudden things: a cyclist, a sign at head height, a bin.
Without it: the closest metre, where collisions actually happen, goes unwatched.
The AirPods
ONE VOICE
- Beeps from the direction of each obstacle; a beacon toward the next turn.
- Speech that knows what matters most and never says it late.
Without it: two cameras and nobody telling you anything.
The beep is
the interface.
Speech is slow. Two words take most of a second, and the person stepping off the curb doesn't have one. So Kite saves its voice for decisions and lets sound carry the distance.
Each object gets a time-to-contact, worked out from how fast the gap is closing. The rate of the beeps follows it, from one a second at the edge of the warning zone to almost nine a second just before contact. Your ears read rhythm faster than any sentence.
Faster means closer. That's the whole manual.
Just say it.
Press an AirPod once and talk. Simple requests run on the laptop with no internet at all; everything else goes to Claude, which sees what Kite sees. These are real replies from the earpiece.
Kite, take me to my usual coffee shop.
Okay, heading to Blue Bottle Coffee.
Bring me to the end of the street.
Okay, I'll guide you to the end of the street.
What's around me?
A person slightly left, 2 meters; a car at 2 o'clock, 8 meters.
Remember this place as the bench.
Saved. I'll remember this place as the bench.
Built to fail safe.
The person using this can't see the screen when something breaks. So every failure has a sound, and nothing smart is allowed to make Kite less careful.
01
Safety never waits for AI
Beeps, "Stop." and "Don't walk" come from plain code that runs first. Claude can tell you to wait. It can never clear a stop or send you on.
02
Late is worse than silent
Every instruction has an expiry. "Turn left now" said three seconds late is dropped, not spoken.
03
"Walk" has to be sure
"Don't walk" plays the instant it's seen. "Walk signal is on" waits until the signal holds steady, so one bad frame can't send you into traffic.
04
Losing sight is announced
If the camera stops sending, the beeps stop and Kite says "Obstacle detection paused." Stale beeps would be worse than none.
05
Urgent goes first
A stop warning cuts off anything less important mid-word, then the interrupted directions come back if they still apply.
06
It keeps working offline
Every safety phrase is recorded in advance. With no internet, Kite still beeps, still says "Stop.", and falls back to a built-in voice.
~0.1 sfrom decision to the first spoken word, streamed from Cartesia
3–12 msto look up a saved place or hazard in Moss
0audio glitches with every CPU core saturated, thanks to a Rust audio engine
168automated tests, including 20,000 malformed messages thrown at it
What each piece actually does.
No logos for decoration. Pull any one of these out and something the walker relies on goes quiet.
Cartesia
THE VOICE, BOTH WAYS
- Sonic streams speech over one open connection: the first word arrives in about a tenth of a second.
- Ink transcribes the walker as they talk, ready the moment they stop.
Without it: a robot voice that arrives a second too late.
Claude
THE JUDGMENT
- Reads the scene and answers "what's around me?" from what Kite actually sees.
- Makes the careful calls ("wait, the crossing is crowded") and is only allowed to make Kite more cautious.
Without it: beeps with no one to explain them.
SpacetimeDB
ONE LIVE WORLD
- Every device on the walk and the caregiver's dashboard share one live state.
- The drone laptop, the earpiece and the dashboard stay in sync without talking to each other directly.
Without it: three machines, three versions of the truth.
Moss
MEMORY OF YOUR WORLD
- "My usual coffee shop" means something: saved places and hazards, searched in milliseconds.
- Arrive at a corner you've marked and Kite reminds you the curb there is tall.
Without it: every walk starts from zero.
Google Maps
WHERE TO GO
- Walking routes become spoken turns and the beacon's direction.
- "The nearest pharmacy" is found by distance from where you're standing.
Without it: you can avoid the street, but not get anywhere.