Building AI Tools for a Photography Studio (No, It Was Not What I Expected)

When I took on the technical head role at KS Photography Station, I expected to manage computers, set up storage, keep the studio software running. I did not expect to build a machine learning pipeline for culling wedding photos. But that is exactly the project that ended up mattering most, and it surprised me at every step.

The problem nobody outside the industry knows

Wedding photography is brutal in a way people do not see. A single wedding can produce two thousand photos, often more. Before any editing happens, someone has to cull. Go through every shot and decide: keep or discard. Eyes closed, out of focus, the same moment shot eight times, pick the best one. This is hours of mind-numbing, high-stakes clicking, and it has to happen for every event.

The glamorous part of photography is the shutter click. The unglamorous part is the human sitting at a screen for hours deciding which clicks were worth it.

What I built

An AI-assisted culling pipeline. Not a magic button that does the photographer's job, that is not how it works and anyone who claims it does is selling something. What it does is the first brutal pass. It flags the obvious rejects: shots that are clearly out of focus, faces with closed eyes, near-duplicate frames where it can group the bursts so the human only compares within a group instead of across the whole pile.

The pipeline scores and sorts so that when the photographer sits down, the worst is already filtered and the duplicates are grouped. The human still makes every real artistic decision. The machine just removes the part that was never artistic to begin with.

• Focus and sharpness detection to flag the obvious technical rejects.
• Face and eye state checks to surface the closed-eye shots in group photos.
• Near-duplicate grouping so the burst of fifteen near-identical frames becomes one decision, not fifteen.

2000 photos, four hours to twenty minutes

The number that made everyone pay attention: a culling pass that used to eat around four hours came down to roughly twenty minutes of human review on top of the automated pass. Not because the machine replaced judgment, but because it removed the noise before the judgment started. The photographer's attention went to the photos that actually deserved a decision.

The skeptical photographer

Here is my favorite part. The lead photographer was deeply skeptical. And he should have been. His instinct was that no algorithm understands what makes a wedding photo special, the emotion, the moment, the thing you cannot put in a loss function. He was completely right about that, and I told him so.

So I never sold it as a replacement for his eye. I sold it as a way to never spend his eye on a blurry throwaway again. After the first real event, where he got hours of his life back without giving up a single creative call, he flipped. He became the biggest advocate for it in the studio. The skeptic turned into the person explaining it to everyone else.

What it taught me

That the best AI tools do not try to do the human's job. They clear the runway so the human can do the part only a human can do. The photographer never wanted to spend four hours rejecting blurry shots. He wanted to spend that time on the photos that mattered. I did not replace his craft. I gave him more room for it. That is the kind of AI work I want to keep doing.