Starting a face recognition company, with eyes open

I’ve co-founded a company. It’s called Amanda AI and it does face recognition. I’ll stay involved at the studio as well, but this is where most of my energy is going now.

I know how the phrase “face recognition company” sounds. A few months ago I wrote about deepfakes and consent. Face technology is getting cheap and powerful at the same time, and a lot of what’s being built with it is creepy. So I want to write down why we’re doing this and how, partly so I can hold myself to it later.

The first product is conference check-in. At a big event, thousands of people arrive in the same hour and stand in line to show a badge or a QR code to someone with a printer. We think attendees who opt in should be able to walk up, look at an iPad, and have their badge print. The line gets shorter and the organizers get their morning back.

We picked this application on purpose, because it’s bounded in several ways that matter. People opt in, and they know exactly what they’re opting into, usually with a photo they submit themselves at registration. The set of faces to match against is small and known: the people registered for this event, not everyone in the world. The data has a natural end date, since the event is over in a few days and there’s no reason to keep anything. And the result is easy to measure. You can count minutes in line.

The other decision is to run recognition on the device instead of in the cloud. Phones and iPads now have enough compute to run a good face model locally, and Apple’s new chips have dedicated neural hardware. On-device matching means faces don’t have to travel to our servers to be compared, it works when the venue Wi-Fi is overloaded (which it always is), and it’s fast enough that nobody waits on a network round trip. It’s harder to engineer, since you’re squeezing models onto constrained hardware and updating them across many devices, but that’s the kind of problem I’ve been working on since my thesis.

The lines I don’t want us to cross: no matching against people who didn’t opt in, no building databases of faces beyond what an event needs, and no selling to anyone whose goal is identifying strangers. The industry is heading toward surveillance uses fast, especially in some markets, and there will be money there. I’d rather build the version of this technology that people are happy to use.

We’ll see if the market agrees. Wish us luck.