I was able to build this using that method:

https://c3pp7f.csb.app/

It uses 1) A decrypted json file for the api key. 2). coco-ssd to find the objects in an image. 3) Gemini AI to describe the located objects.

It is 100% AI code built by recycling around and around from Google AI Studio to Claude/other AI.

When I am 100% happy (possibly never) I might then build with v0.dev.

e.g. like this:

https://0cgnttfj6gdnahym.vercel.app/

I have another Elevenlabs "better voice" version but I can't afford to let other people use my credits.