01A game with the camera
It was a game. I'd point the camera at someone at home, or at the dog, and ask it to memorize them. The first version landed on 2 June: it saved the photo, the name I said, and a written description of the features (“dark hair, glasses”).
02What it was really measuring
On 11 September I sat down to find out why it failed, and the problem wasn't the model. To recognize someone, the agent compared written descriptions plus a generic fingerprint of the image. Ask anyone to describe two brothers' faces in words and you'll get two nearly identical paragraphs: same hair, same glasses, same smile. A system like that can tell a person from a beach. It can't tell two people who look alike apart, and that's the only case where recognizing someone means anything. It was measuring something else: how alike two sentences were.
Underneath was a worse gap: “who is this?” didn't exist as a question. You could only search a phrase against the saved photos, which works for “show me someone with glasses” but not for finding out who's pictured, so the agent looked at the image and guessed. In my own records, the same person appeared three times.
I'm not a face-recognition specialist; I do know how to ask what a system is really measuring.
03That night
I decided at 21:51 and by 22:45 it was live: a facial recognition service, in the EU, that compares the face, not a description of it. The faces I had already saved were moved over that same night, under three rules.
The name doesn't travel: the service gets a meaningless identifier, and the name stays in Saelyx's database, where you can delete it. Each user has their own index, so deleting an account is a single call. I don't ask for age, gender or emotion either: the service can estimate them, but you don't need them to know whether it's the same person. And “I don't know her” counts as a complete answer. The match threshold is high on purpose: telling someone “this is your sister” when it isn't costs more than saying “I don't know her.”
04Two failures that weren't the model's
On 12 September, testing on the iPhone with “Remember faces” off and nobody introduced, the agent “recognized” me by a red cap and a white T-shirt. It hadn't called the service: it compared the scene with earlier photos and presented that as recognition. Now, when it can't know, it says so and explains why.
The same day, on my partner's iPhone, another failure turned up: the same face saved twice under two names. Every frame matched both records at 99.99%, and the agent insisted it was one or the other. The service was right, it was the same face; the memory was wrong. Since 13 September a face that already has a record isn't saved again, and when two names match the agent doesn't pick one: it says the face looks like both, and asks.
05With permission, and no guessing
There's a reason for all that care: a face is biometric data, a special category under European data protection law. “Remember faces” is off until you turn it on, and withdrawing it deletes every saved face. It only recognizes people you've introduced; it doesn't name anyone else or describe their features. Saelyx is for people 18 and over, and whoever shows another person's face is responsible for having a legitimate basis to do so. The details are in “Your memory in Saelyx.”
In a meeting, where nobody has been introduced, the agent reads, on the Mac, the name the video call writes under each tile: reading text isn't biometrics. The face stays as a complement, and only with your consent.
I'm writing this on 6 October 2026 and I don't have an accuracy rate to show you; I'd rather not make one up. I do know which way I want it to fail: when in doubt, Saelyx doesn't know anyone.
06Frequently asked questions
How does Saelyx recognize a person?
With a facial recognition service, in the EU, that compares the face rather than a description. It only searches among the people you've introduced; if nothing matches, the agent says it doesn't know the person.
Is it on by default?
No. “Remember faces” is off until you turn it on, it's a paid feature, and you can withdraw it any time: the index entry, the photo and the record are deleted.
Does it recognize anyone who appears on camera?
No. Only people you've introduced; it doesn't name anyone else or describe their features, and the service isn't asked for age, gender or emotion.
— Adianny
Senior DevOps engineer. I've spent about four years working with machine learning and MLOps, now as a tech lead; Saelyx is my side project.