Thousands of faces in a hotel lobby
Our first live conference check-ins. Wait times dropped a lot. The model was the part that worked from day one. Lighting, queues and printers were not.
We’ve now run face check-in at our first real conferences, with attendees who opted in walking up to iPads and getting their badges printed without showing anything. I’ve spent a lot of this summer in hotel ballrooms at 7 a.m., and I learned more there than in six months of testing in our office.
The headline result is good. For people who used face check-in, the time from walking up to holding a badge dropped to a few seconds, and lines were much shorter than at the traditional desks next to us. Organizers noticed, which is what matters for a startup.
What surprised me was where the problems were.
The recognition model was the least of our worries. We’d spent a year getting it accurate and fast on device, and it did its job. When it failed, it was almost always because it never got a usable image.
Lighting was the first real enemy. Hotel lobbies are designed to look nice, not to light faces evenly. We had one venue with a huge window behind the check-in area, so every attendee was a silhouette against the morning sun. Another had warm spotlights that put hard shadows under everyone’s eyes. We ended up carrying our own lights and choosing kiosk positions ourselves on the first day, which no amount of model training could have replaced.
Queues were the second. People don’t stand where you expect. They walk up at an angle, talk to a colleague, look at their phone, wear sunglasses pushed up on their heads. The system has to decide when someone is actually trying to check in versus passing by, and it has to do it without making anyone feel watched. Small interface details, like a clear prompt and a mirror-like preview, made a bigger difference than a percentage point of accuracy.
Third, and least glamorous: printers. The badge printer is the slowest part of the whole flow, and at one event two of them jammed in the first twenty minutes. The attendee doesn’t care whose fault that is. For them, the face check-in “didn’t work.”
I think this is true of most applied AI. The model is one piece, and often it’s the piece you can control best. What decides whether the product works is the whole chain around it: the environment the sensor sees, how people actually behave, and the boring physical hardware at the end. For the next round we’re spending more time on the chain than the model.