The same prompt, the same model, the same boxes as every case on this site — reading your camera instead of a generated clip. Best on a phone, held up to a windscreen.
Sign in to report what it finds.
A report is emailed to you, with the marked-up still attached, and kept on your incidents page. So it needs an inbox to go to. Signing in is Google, and takes one tap.
Camera off.
A look every 2.5 seconds, each one a call
to gemini-3.7-flash. Up to 6 of them are in the air
at once, so a slow answer holds up the picture rather than the next question.
Vertex throttles at roughly twenty to thirty a minute, so this is a short demo
and the counter tells you where you are. When it finds something it stops
looking — a hazard is one object, and more calls on it buy nothing.
Only what you report is kept. Every frame goes to the model and is dropped. A find is held for 15 seconds and then discarded too — if you do not press the button, nothing is uploaded, nothing is written down, and no mail is sent.
Press it and one thing is kept: that frame, with the box drawn into it, stored against your account and emailed to you with the finished report. It works out which agency owns that stretch and the mail says so — but it does not write to them. Forwarding it stays your decision, and your hand. Share it yourself is the old path, and still stores nothing.