Here’s the plan I was sure would work: point an AI at a folder of thousands of mixed photos, give it a few pictures of my daughter, and let it pull out every photo of her automatically. No more manual sorting.

It didn’t work. And the reason it didn’t work is interesting enough that I want to tell it straight, including the part where I was wrong.

The Idea

Face recognition is a solved problem, right? Phones do it. The plan was:

  • Give the model a handful of reference photos of my daughter.
  • Have it scan the big pile and score each photo for “is this her?”
  • Move the likely matches to a separate folder, and let me correct it so it got better over time.

I even planned for it to run slowly on a low-spec PC — I didn’t need speed, just a model small enough to run at home, learning as it went. My daughter was born in 2024, so I figured I had years for it to improve.

💬 Prompt that worked “I have one folder with thousands of mixed photos and I want to pull out only the ones of my daughter into a second folder. Can I train a model on a few reference photos of her, move the likely matches, and feed my corrections back so it improves over time? The PC is low-spec and slow is fine. Plan it as a sequence of steps I can run in order, since the source folder is huge.”

That prompt got me a sensible plan and working code. The plan was fine. Reality was the problem.

Why It Failed: A Baby Isn’t One Face

Face recognition assumes a person looks roughly like themselves over time. Adults do. A baby does not.

My daughter at one month, six months, one year, and two years is, to a model, almost four different people. The soft newborn face, the chubby six-month face, the toddler face — they don’t match a single reference the way an adult’s face matches across a decade. Every few months, the model lost her.

I could have fought this: more reference photos per age, re-training at each stage, thresholds tuned by month. But think about what I was actually building — a way to save time sorting photos. By the time I babysat a model through every stage of my daughter’s changing face, I’d have spent far more effort than just… looking at the photos myself.

The “Worse” Approach That Won

So I dropped the AI and built the plain thing: a browser gallery where I look at the day’s files and tap the keepers. (That’s the picker from the last post.)

On paper it’s less clever. In practice it’s better here, for reasons I didn’t appreciate until I tried the smart version:

  • It’s never wrong. I know my daughter’s face at every age instantly. No model does.
  • It’s fast enough. A day’s new files take about a minute. The AI would need to process thousands and still be checked.
  • It costs nothing. No model to run, no API to call, no per-image charge. Just my eyes.
  • I want to look anyway. The “chore” is a minute a day of seeing my daughter’s life. Automating it away would remove something I actually value.

What I Took From It

I’m not anti-AI — this whole blog is about building with it. But this was a good reminder that “add AI” is not automatically the answer, especially for personal software where you’re the only user and the human judgment is cheap and accurate.

The honest scoreboard: the fancy approach cost me a few evenings and produced nothing usable. The simple approach cost me a minute a day and produced the entire archive. Sometimes the boring tool is the right tool, and the useful move is to notice that early instead of defending the clever one.

The next posts go back to the parts I did automate well — starting with why my daughter’s videos wouldn’t even play until I converted them.