Stop pitching it. Build it.

I've sat through enough planning meetings in retail to know the question that never gets a good answer: why didn't that product move? Sales data tells you what sold. It never tells you what was seen and passed over.

The obvious fix — face-tracking analytics — answers the question and loses the customer. I wasn't willing to design that. So the idea became: measure attention on products, never on people. People are dots. Shelves hold the data.

The deck I didn't make

A year ago this would have been twelve slides and a request for engineering time. Instead, I built it: a three.js point-cloud digital twin of a convenience store, simulated shoppers walking the aisles, a live heatmap counting attention by shelf zone. One HTML file, no build step, an end-to-end test suite driving it in headless Chrome.

A working thing you can click ends a debate that a mockup only starts.

When someone asks "but what about privacy?", the demo answers it structurally: there is no video to leak, no face to store, no individual path to subpoena. The system keeps seconds of attention per zone and nothing else. That's not a policy promise — it's the architecture.

What AI actually changed

AI didn't design ShelfSight. It collapsed the distance between the idea and the artifact. The framing, the privacy rule, the three jobs the product had to do — that's still design judgment. But the gap between "I think this should exist" and "here, try it" went from a quarter to a weekend.

That changes what a designer is for. The scarce skill isn't producing screens anymore; it's knowing which thing to build, what rule it must never break, and having the taste to say when it's right. Then building it, because you can now.