The honest version
I’m Stefan Ćirović — an AI-native product builder from Belgrade, Serbia. I handle design and strategy, and I ship production software by directing AI agent teams: websites, web apps, games, and mobile products, end to end. Self-taught, with 10+ years in visual design and a decade of running operations for one of Serbia’s leading beauty companies.
I don’t write code.
I need you to know that up front, because everything else on this site makes more sense once you do. I’m a designer by trade and self-taught in everything — my formal education ends with a technical high school in Belgrade. No CS degree, no bootcamp, no certificates. What I have instead is a lifetime of taking things apart to see how they work, and — since AI agents became capable enough to build production software — a method for directing them that’s stricter than anything I’d get away with in a human team.
The results are on the Work page: a real-time logic duel game live in 13 languages, a painter’s gallery with a real-time auction engine, a phone app that translates your voice mid-call, a client site that went from brief to production in a single day. I designed for a decade before any of this — print campaigns for international beauty brands, logos, visual identities — so the things I ship don’t just work; they look like someone cared.
How I work — the loop
“Built with AI” in 2026 means anything from a weekend toy to production software. The difference is the process, and mine runs the same loop every time — no shortcuts.
Every project starts with conversation, not code. We map the business problem, the users, the money, and the risks — and write it down. My rule, from my own chat logs: “I don’t want to head in the wrong direction and waste time on fixes — or worse, start over.”
Features get written specifications. Anything high-stakes — payments, auctions, anything that touches money — gets adversarial research first: what other people’s audits found, what went wrong, what we won’t repeat.
Every meaningful choice gets a numbered entry: what we decided, when, and why. When I paused my translation app, the decision book held 27 product decisions — I can tell you the reasoning behind every one.
Small, closed units of work. No thousand-line surprises.
After every segment, a separate AI agent with no memory of writing the code audits it. Findings get classified and re-audited until blockers and majors hit zero. My game SYNAPSE went through six numbered audit rounds.
Which Serbian voice for a translation app? Picked by listening, not by spec sheet. Speech latency? Measured through the real API. A bug we couldn’t explain? A five-line experiment settled it in one run (that story here).
Agents work autonomously — including overnight shifts — but production is a ritual: nothing deploys until I’ve personally tested it on real devices. My eyes are the last gate.
What I do myself vs. what I delegate
| I do | My agents do |
|---|---|
| Product vision, business model, pricing | All code, all architecture |
| Design direction & pixel-level QA | Research, documentation, tests |
| Naming, brand, copy direction | Audits (always fresh context) |
| Manual testing on real devices | Automated test suites |
| Talking to humans | Talking to APIs |
A solo founder with AI agents isn’t a smaller team — it’s a different shape of team. No coordination tax, no knowledge silos, one person holding the entire product in their head, and agents that never get tired of writing tests. The discipline isn’t overhead; it’s what makes the speed safe.
The record
My own practice, under my own name: one person, a team of AI agents, and a shipping streak you can verify page by page.
A design and marketing agency serving clients end to end. The agency is no longer active — it wound down in 2025, and its lessons became the blueprint for what I do now.
Operations & Creative Manager for the beauty company behind Paul Mitchell in Serbia: print and digital campaigns, a full rebrand, IT systems, event logistics, sales strategy — a collaboration that still continues today.
IT support, system migrations, staff training. The self-taught kid gets a badge.
The tools I actually reach for — Photoshop, the agent stack, Cloudflare edge, and everything else in the kit — have their own page: the Toolbox.
What I believe
In my game, the server is the sole referee — round times are measured server-side, and abandoning a match can never profit the deserter. May the better player win.
My translation app doesn’t record conversations — not by policy, but by design: the audio has nowhere to be stored.
When my team debates, we build the experiment that settles it. It’s cheaper than being confidently wrong.
Even a one-day website for a local plumber gets structured data, four languages, and an admin panel.
Off the clock
Dota 2 (inspired by the pros, humbled by the ladder), techno and trap, and digital art. Father, husband, Belgrade native.