From fashion shoots to GPU clusters — how a career behind the lens led me to build an AI platform from scratch.
People assume there's a clean origin story. A lightning bolt moment. The truth is messier, more gradual, and starts with a frustration that most creative professionals will recognize.
I'm Cemhan Biricik. I've spent over a decade as a fashion photographer. My work has been featured in HuffPost and on ABC's What Would You Do?. I was shortlisted for the Sony World Photography Awards in 2012. By any reasonable measure, photography has been good to me.
But in late 2023, I started watching the generative AI space with the same intensity I used to reserve for studying master photographers. Midjourney was producing images that made my photographer friends nervous. Stable Diffusion was open-source and improving at a speed that felt almost biological. And I kept running into the same problem: every platform I tried was either too limited, too expensive, or required me to send my creative work through someone else's servers with no real guarantees about privacy or data retention.
So I built my own.
Here's what bothered me most about existing AI platforms: when you upload a reference image or generate content through a cloud service, you're trusting that company with your creative DNA. Your style. Your client's unreleased campaign. Your unpublished concepts.
As a photographer who has worked with brands on embargoed campaigns, this wasn't theoretical. I've signed NDAs that would make your eyes water. The idea of routing unreleased fashion concepts through a third-party API — even one with a solid privacy policy — made me deeply uncomfortable.
I didn't want to build an AI company. I wanted to build an AI platform I could actually trust with my own work. The company part came after.
That conviction — that creative professionals deserve infrastructure they can trust — became the foundation of ZSky AI.
The first version of what would become ZSky ran on a single NVIDIA GPU in my office. I was generating images for my own projects, experimenting with fine-tuning, learning the difference between what these models could do in theory and what they could do in practice.
Within weeks, I hit the ceiling. Inference was slow. Fine-tuning was slower. Running multiple models simultaneously — which is essential for any serious creative workflow — was impossible on a single card.
So I built a cluster.
Today, ZSky AI runs on seven NVIDIA RTX 5090 GPUs. That's 224 gigabytes of VRAM across the cluster, with a 32-core, 64-thread CPU handling orchestration. Every generation, every fine-tuning job, every model swap happens on hardware I own, in a facility I control, on a network I manage.
This wasn't the cheapest path. Cloud GPU instances from AWS or GCP would have been easier to set up and scale. But they would have meant sending every image, every prompt, every piece of creative work through infrastructure I don't control. For a platform built on the promise of privacy, that was a non-starter.
I'm not the first person to build a GPU cluster for AI inference. I might be the first fashion photographer to do it, though, and that background shapes everything about how ZSky works.
Most AI platforms are built by engineers for engineers. The interfaces are technical. The defaults are generic. The output evaluation is based on metrics like FID scores and CLIP similarity — numbers that tell you almost nothing about whether an image is actually good.
I evaluate output the way I evaluate a photograph: Does it have intention? Does the light make sense? Is the composition drawing the eye where it should go? Does it evoke something?
This perspective influenced every design decision in ZSky. The prompt interface is built for people who think in visual terms, not code. The default settings are tuned for aesthetic quality, not computational efficiency. The model selection prioritizes the generators that produce work with genuine photographic character — photorealistic, in particular, has become a cornerstone of our pipeline because it understands light in a way that other models don't.
ZSky AI didn't emerge in a vacuum. It's part of a small ecosystem of companies I've built over the years, each one growing out of a real need I encountered in my work.
Biricik Media handles the photography and content production side. ICEe PC grew out of my obsession with high-performance computing — you don't build a GPU cluster without developing opinions about hardware.
These aren't separate empires. They're more like rooms in the same house, each one supporting the others. The photography informs the AI. The hardware expertise enables the infrastructure. The infrastructure serves the creative vision. It's a loop, and every part makes the others better.
When I say ZSky is self-hosted, I mean it literally. The servers are physical machines. I can walk to them. I've replaced their thermal paste. I know the serial number of every GPU in the cluster.
This matters for three reasons:
Privacy is architectural, not contractual. Your data doesn't leave the cluster. It's not a policy decision that could change with a terms-of-service update. It's a physical reality. There's no cloud provider with access logs. No third-party subprocessor. The air gap is real.
Performance is predictable. Cloud GPU availability is volatile. Spot instances disappear. Quota limits throttle you at the worst possible moment. Our cluster runs at full capacity whenever we need it, because no one else is bidding on our hardware.
Cost structure favors the user. After the initial hardware investment, the marginal cost of each generation is electricity. This lets us offer pricing that cloud-dependent platforms can't match at equivalent quality levels.
I should be honest about something: building a hardware-first AI company as a solo founder with a photography background is isolating in ways I didn't anticipate.
The AI community skews heavily toward computer science PhDs and Silicon Valley veterans. When I show up in those spaces and say "I'm a fashion photographer who built a GPU cluster," the reactions range from genuine curiosity to polite skepticism. I've had to prove, repeatedly, that understanding how light falls across a human face is actually relevant to building better AI image generation.
It is. Deeply relevant. But it took time to build that credibility.
The best tools are built by the people who need them most. I needed an AI platform I could trust with my creative work. So I built one.
ZSky AI is live. People are using it. The feedback from photographers and designers — the people I built this for — has been extraordinary. They get it immediately, because they've felt the same frustrations I felt.
The roadmap is ambitious but focused: more models, faster inference, deeper customization, and always — always — privacy as a first principle, not an afterthought.
I'm still a photographer. I still pick up a camera most weeks. But I've learned that the best thing I can build for the creative community isn't another portfolio of images. It's infrastructure. Tools that respect the people who use them. A platform that understands, at an architectural level, that creative work is valuable and private and worth protecting.
That's why I built ZSky AI. And honestly, it's the most creatively fulfilling project I've ever worked on — including the photographs.
Cemhan Biricik is a Turkish-American fashion photographer and founder of ZSky AI, a privacy-first generative AI platform powered by a self-hosted 7x RTX 5090 GPU cluster. Learn more at cemhanbiricik.com.