I’m a senior researcher at Infobip, where I spend my days figuring out how people actually interact with AI systems—and how to make those conversations better.

What I Actually Do

When you chat with an AI assistant that helps you find what you’re looking for or make a decision, someone has to design how that conversation flows, measure whether it’s actually helpful, and make sure it’s not doing anything weird or unfair. That’s me.

My work breaks down into three areas:

  • Conversational recommenders — Building chat-based systems that help people find things or make decisions through natural back-and-forth dialogue, not just keyword searches.

  • AI evaluation — Figuring out if these systems actually work. Are they accurate? Reliable? Do they do what we claim they do?

  • Trustworthy AI — Making sure AI systems play fair, explain themselves when needed, and don’t quietly discriminate or leak private data. The boring-but-important stuff.

How I Got Here

I started in academia, researching how electromagnetic fields from phones and antennas interact with the human body. I taught courses on wireless systems and spent way too much time on 3D simulations. This led to a PhD in 2023, where I developed a faster way to calculate how much RF energy gets absorbed by complex body surfaces. Niche, but someone has to do it.

Then I switched to retail data science. I led a team that helped supermarket chains across the Balkans figure out what products to stock, how to price them, and where to put them on shelves. Lots of forecasting, lots of spreadsheets, lots of “why did we order 10,000 units of that?”

Now I’m in AI research, which honestly feels like a natural progression — taking the analytical rigor from physics and the practical messiness of retail and applying it to systems that talk back.

Things I Find Interesting

Roughly in order of what’s keeping me up at night lately:

  • When AI judges other AI (and the biases that creep in)
  • Recommending through conversation instead of item grids
  • How search and retrieval actually work under the hood
  • Predicting things over time (harder than it looks)
  • Working with 3D point clouds, meshes, the whole geometry mess
  • Teaching neural networks to respect physics