The unpopular route
The default story right now is: train a model, wrap it, sell it as B2B SaaS, scale it to the moon. I love that game. But I wanted to play a different one.
I built a small A.I. business for my family to work at. It’s local, it doesn’t scale, and it requires people to run it day-to-day. My wife is at the kiosk. My mom is at the kiosk. The whole point is that AI here creates jobs instead of replacing them.
That framing changed every product decision. Instead of asking “how do we remove humans from this loop?”, the question became “how do we put humans at the center of an experience that AI alone couldn’t deliver?”
What it actually is
AI Photo Lab is, on paper, very simple: a superpowered photo booth in a mall. You walk up, take a picture, pick from a list of popular and trendy themes, watch the AI reimagine the world around you, and walk away with a printed 4×6 in your hand.
But the customer in front of the kiosk is rarely a tech person. It’s a family with three kids shopping for shoes. It’s a grandmother who came to the food court. It’s teenagers killing time before a movie. None of them are going to read a prompt guide. They have about fifteen seconds of patience before they walk away.
Designing for that audience is the actual product.
UX for the non-technical, in public
A mall kiosk is a brutal user-research environment. The interface has to work for a five-year-old and a seventy-five-year-old. There’s ambient noise, mall lighting, and zero patience for friction. Some constraints I committed to early:
- No app, no account, no prompt. If a step requires someone to download or sign in, we’ve already lost half the foot traffic.
- Big tap targets, plain language. No jargon, no “seed,” no “strength,” no “negative prompt.” Just pictures and words.
- The screen always knows what to do next. One primary action at a time. Everything else is secondary.
- Themes are visual, not textual. You don’t describe what you want. You point at a picture of it.
- Faces are sacred. If the AI changes someone’s face, we’ve broken the experience. People came to see themselves in another world — not a stranger.
The last one is the hardest engineering constraint and the one I’m proudest of. The whole pipeline is built around preserving identity while letting everything else change.
The portrait pipeline
A short tour of the parts that actually matter. The API call is the easy part. The interesting work is everything around it.
1) Models — latest & greatest, always
The image-generation space is moving so fast that committing to a single model is a mistake. I treat the model layer as a swappable component. Whatever wins this quarter is what runs in the booth next quarter. No legacy stack to defend.
2) Prompts — engineered, not improvised
Each theme has its own carefully crafted, tested, and polished prompt behind it. They’re not one-liners. They’re structured templates that lock down composition, framing, and identity, and then leave room for the world around the subject to change.
// Pseudocode — every theme is built from these layers.
const prompt = compile({
identity: "preserve subject's face, age, skin tone, hair, expression",
framing: "match original pose, crop and gaze; no extra figures",
world: THEME.description, // anime / space / superhero / ...
style: THEME.styleDirectives, // palette, lens, lighting
guardrails: "family friendly, no text artifacts, no logos, no warping",
});
The boring layers (identity, framing, guardrails) are constant across every theme. Only the
world and style change. That’s why faces stay yours and the worlds don’t.
3) Themes — a living catalog
Themes are content, not code. They’re curated like a magazine: pop-culture moments, seasonal drops, viral aesthetics. New looks ship regularly so the kiosk never feels stale and there’s always a reason to come back. Customers don’t care about model versions — they care that the “anime” option this month looks like this month, not last year.
4) Output — fast enough to feel magic
Images come back in roughly ten seconds each. Anything longer and you’ve lost the room. Anything faster doesn’t feel like AI is “doing something special.” That window is part of the product.
5) Privacy — built in, not bolted on
No permanent storage of customer photos. No model training on their data. No human review pipeline. Photos are processed and gone. This isn’t a marketing claim, it’s the architecture: there’s literally nowhere a face lingers once the session is over.
From a digital moment to something you can hold
Almost every AI experience right now ends inside a screen. That’s where the magic dies. A printed 4×6 in someone’s hand changes the emotional register completely — it becomes a keepsake, a fridge magnet, a gift, a thing to show their cousin in another city.
The print path has its own engineering: color profiles tuned for the dye-sub printer in the booth, generation resolution sized for clean print output, and a queue that survives the worst-case scenario where ten kids tap the same button at the same time.
What I’ve learned shipping a non-scaling AI product
Constraints make AI products good
Same lesson I keep relearning. Removing user choice (no prompt, no parameters, no “strength” slider) is what made the experience accessible. Every removed knob is a customer who didn’t walk away.
Theming is a content problem, not an AI problem
The model is the engine. The themes are the menu. We win by being the people with the best menu — the most current, the most photogenic, the most fun — not by having the most clever inference setup.
Local doesn’t mean small thinking
One kiosk, one mall, one staffed booth. But every operational decision (queueing, payment, content moderation, printing reliability, customer support) has to actually work in the wild. There’s no “we’ll fix it in v2” when a kid is crying because the printer jammed.
AI can create jobs
The most surprising thing about this project, honestly, is how good it feels. The AI doesn’t replace anyone here — it gives my family a real, hands-on business to run. People walk up, smile, and walk away with art they made together. That part isn’t a tech demo. That part is the whole point.
Come say hi (and grab a free print)
In between all the same stores people are used to, families and older folks and non-technical customers are bumping into innovation made simple enough for them to try, have fun, and be amazed. That’s the most rewarding loop I’ve built in years.
If you’re in Seattle, you can find us at AI Photo Lab inside Westfield Southcenter in Tukwila. My wife or my mom will probably be at the kiosk. Tell them you know me — print is on me. 🫡