Hello, world 👋
My name is Krsto. Yes, four consonants in a row. I’m an engineer and researcher interested in artificial intelligence.
As an engineer, I develop AI agents that solve complex, real-world tasks. I currently work at EPAM Systems.
As a researcher, I’m trying to understand how self-organisation gives rise to adaptive behaviour, including learning algorithms. See below for my latest work on incorporating plasticity into neural cellular automata.
Outside of work and research, I enjoy outdoor sports like triathlon, trail running, and open water swimming. I’m also into video games, particularly indie games.
I'm originally from Montenegro, where I'm currently based. I've also lived in Belgium, Netherlands, India, and Switzerland. I speak Serbo-Croatian, English, and Italian.
Elsewhere
Favourite work
-
arXiv:2609.29281
We introduce neural cellular automata that adapt online to new tasks without computing gradients at adaptation time. Each cell maintains its own fast plastic state, updated from prediction error, so task adaptation emerges through spatially distributed changes in the recurrent dynamics while the slow parameters remain fixed.
-
IASEAI 2026
We audit a Dutch public-sector risk-profiling algorithm used to assess fraud risk among college students, showing how unsupervised bias-detection methods can identify potentially discriminatory patterns even without access to protected demographic attributes. We find patterns consistent with previously documented disparities affecting students with non-European migration backgrounds, while emphasizing that the method is intended as a screening tool for further human and legal review rather than as proof of discrimination. We also introduce an open-source Python package that makes the proposed bias-detection approach available for use on other datasets.
-
Meta-Learning for Recalibration of EMG-Based Upper Limb Prostheses
Automated Machine Learning (AutoML) Workshop & Lifelong Learning (LifelongML) Workshop, ICML 2020
We show that meta-learning is a promising approach for efficiently recalibrating EMG-based upper-limb prostheses when muscle signals change between recording sessions. Our results demonstrate that meta-learning can outperform conventional pretraining and training-from-scratch approaches while requiring only a small amount of recalibration data. We also show that recalibration remains possible even when examples of some movement types are unavailable in the new session.