The first goal was modest: I wanted wired audio from my devices that do USB-C audio but not Android Auto or CarPlay, with Bluetooth and Wi-Fi switched off. That's not where it ended.

A Raspberry Pi now pretends to be a phone toward the car, an ESP32 pretends to be a headset toward my phone, and a small service on my NAS fetches a fresh Android Auto certificate every week so the car keeps trusting it. The car screen shows cover art, the queue and turn-by-turn directions from Google Maps, and the steering wheel buttons work.

The first real test in the car passed 9 of 9 checks. First try. That is not how hardware usually goes. The AI also did the hardware side: wiring diagrams, a two-part case for the 3D printer generated from a script, and a circuit board with the files ready to send to the factory.

My mum has dementia. Her tablet shows big text, today and tomorrow, a landscape at the top that follows the weather and the seasons, a big button for the radio, card games you can't lose, and a dog called Bella who is always happy and never needs walking.

The home helpers can cancel or add an appointment with a code and a photo, so the family can see who changed what. Everything is built around dementia guidance: one fixed layout, very big text, pictures instead of words, and nothing you can get wrong.

It also came with its own Android app that takes over a cheap tablet completely: full screen, always on, locked to the calendar, and it starts again by itself if the tablet restarts.

I never trusted the app that came with my cheap NVR. Now the recorder can be banned from the internet completely, and I still have my own way to control it.

With REA, the agent looked inside the vendor's own library to check how the protocol really works, and then built a replacement that only talks to the recorder on my own network. Live view, recordings, two-way talk and faces grouped by person, all running on my NAS.

I think it's the same trick behind ArtCraft, the free Adobe clones that have been all over the feed lately: REA in a clean-room setup, where you first use it to write docs on how the app works, and then rebuild the whole thing from those docs.

Saturday evening it was a grey ball rolling down some grey hills. By midnight it was made of candy, had music, and an evil licorice ball was racing you. By Sunday evening it had ten levels, a level editor and tilt-the-phone controls.

Every model is a Python script, every level is an ASCII map, and a bot plays all the levels after every change, so nothing ships untested.

Then I joked on a LAN party's Facebook event that I might make a multiplayer version. So now I had to. This weekend it got online races for six players, a server on my NAS, and candy that makes you dizzy.

Play it here: vlx42.github.io/candy-marble