Zagreb and Linz
I'm a software engineer. At Ericsson Nikola Tesla I write Java for the test framework that validates Ericsson's 4G and 5G radio network. I also co-founded Vectra XR, a 3D configurator for custom fitness racks and prefab homes, and I'm its CTO.
I finished my bachelor's at TVZ in 2026. My thesis put three strong language models and one weak local model into the same economic games. Now I'm doing the MSc in Artificial Intelligence at JKU Linz.
Asymmetric Society
When language models of very different capability share a task, does the stronger one come out ahead?
Three strong models, Claude Haiku 4.5, GPT-5-mini and Gemini 2.5 Flash, play repeated economic games with one weak model, Llama 3.2 3B, run locally.
With no communication the strong models earn a capability premium. Add a public pledge before each move and the premium inverts: the strong models earn less.
Vectra XR
See it before you build it. A 3D configurator for things people buy before they are made: custom fitness racks and prefab homes.
A buyer configures a rack in real time, changes its colours and attachments, and sees it at full scale in their own room through the phone camera, with no app. For a prefab home, each buyer gets a private 3D walkthrough of their own version, with finishes, layout and live pricing.
anneal
Edit a real crystal structure and watch a machine-learned potential relax it, one step at a time.
kessler
Close approaches between satellites in the public catalogue, and what a small burn does to one.
surrey
A lidar segmentation model trained on Surrey, British Columbia, and run on San Francisco, a city it never saw.
San Francisco's Financial District, coloured by what the model calls each point.
Earlier projects
Now
Before
What I work with
Croatian, native. English, C1.