Personal · AI showcase
Jev Park
A car that finds a parking space and parks in it, in a lot you design. Every judgment on the way, from whether a space is empty to whether to wait for the car backing out, is asked live of Jev, TypeSafe's fast vision-and-text model, several times a second.
What it is
Pick a lot (a café, a post office and a gas station sharing an odd plot, a corner lot, a small plaza, a supermarket), say how busy it is, switch on the traps, and write what the driver wants in their own words. Or type an idea and let Jev set the scene. Press Play.
The lot, the cars, the shoppers and the car's own driving are simulated. The judgment is not: a camera on the car's roof sends frames to Jev, which says what it sees, and Jev's text side plans from what it has been told. A car backing out, a cart left in a space, a motorcycle that makes a space look empty from far off, two pickups over the lines: Jev has to notice, and the page shows what it said, where, and how sure it was.
Where Jev decides
- Eyes, about four times a secondA frame and up to seven close-ups in one request: is each space empty, what kind is it (accessible, electric, compact), how easy to get into; what is happening ahead; which car shows reverse lights; at a junction, the gist of each lane.
- The planner, about once a secondFrom the driver's wish, what the eyes have seen and a log of what has happened: which space, which lane, wait for the car leaving or drive on, and, after a while without a space, whether to settle, wait, try the far end or leave.
- The refereeWhen two cars each stop for the other, Jev says which one backs up.
- Not Jev: the car's own drivingCode steers to the chosen space, stops for people and cars, and brakes for anything in front, as a real car's own systems would. The page says which is which.
Measured
On the lot it opens with, every live run parks, in about 65 seconds and 100 metres of driving, and makes about 260 calls to Jev for five cents. A vision call's round trip from Dallas is about 170 ms, 140 to 165 ms of it at TypeSafe; it carries a frame and up to seven close-ups, about 5,000 tokens. Asked whether a space is empty, Jev agrees with the simulator's truth 83% of the time, and some of the rest are spaces a passer-by stands in front of; asked what kind of space it is, 99%.
Unofficial: not affiliated with or endorsed by TypeSafe. Jev is TypeSafe's model (jev-1.14.0), called through their API by a small server that holds the key and asks only fixed questions; the visitor's browser never sees the key. Not a driving aid.