CroAI – ML Pub Meetup
Zagreb, Croatia
Invited talk on Superhuman Planning with Large Language Models hosted by CroAI.
About Me
I co-developed a Gemini Chess Gem for Gemini—the first language model to reach Grandmaster-level performance in Chess with a human-comparable search budget per move. It integrates search-based planning techniques to enhance multi-step reasoning in games like Chess, Chess960, Connect Four, and Hex. This work was part of my Student Research at Google’s Gemini Post-Training Team, hosted by Eric Malmi and Aliaksei Severyn.
I obtained my PhD at ETH Zurich on safe and scalable multi-agent reinforcement learning and was an Associated Researcher at ETH AI Center, supervised by Prof. Francesco Corman and Prof. Andreas Krause.
Before starting a PhD, I worked as a Machine Learning Researcher at Morgan Stanley and as a Senior Maching Learning Researcher, leading a team of four, in a tech startup, Cantab Predictive Intelligence.
Interests
- Artificial Intelligence
- Reinforcement Learning
- Large Language Models (LLMs)
- Planning and Reasoning (with LLMs)
- Inference Time Methods
- Sequential Decision Making
- Multi-Agent Systems
- Probabilistic Learning
- Safe Learning
- Data-Driven Algorithms
- AI in Board Games
- Mean-Field Control
Education
-
PhD in Artificial Intelligence
ETH Zurich
-
MSc in Mathematical Statistics
University of Zagreb
-
Visiting Student
University of Bielefeld
-
BSc in Mathematics
University of Zagreb
Featured Papers
Mastering Board Games by External and Internal Planning with Language Models
ICML 2025 (Spotlight)
May 23, 2025
Ride-Sourcing Vehicle Rebalancing with Service Accessibility Guarantees via Constrained Mean-Field Reinforcement Learning
arXiv preprint
Apr 1, 2025
Safe Model-Based Multi-Agent Mean-Field Reinforcement Learning
AAMAS 2024 (Oral)
Dec 26, 2023
Recent & Upcoming Talks
ETH Zurich AI Center Associated Researchers Meetup
Zurich, Switzerland
Invited talk on Mastering Board Games With Language Models hosted by ETH AI Center.
ZurichNLP Meetup
Zurich, Switzerland
Invited talk on Mastering Board Games With Language Models hosted by ZurichAI.
Google DeepMind Booth at NeurIPS 2024
Vancouver, Canada
A talk on Mastering Chess With Language Models.
The 23rd International Conference on Autonomous Agents and Multi-Agent Systems (AAMAS 2024)
Auckland, New Zealand
A conference talk on a Safe Model-Based Multi-Agent Mean-Field Reinforcement Learning.
Experience
-
Zurich, Switzerland
Student Researcher
Google –– Gemini Post-Training Team
- Key Contribution: The first LLM that plays chess at the world champion level using human search budget — Gemini Chess Gem.
- Hosts: Eric Malmi and Aliaksei Severyn
- Publication: First co-author of a spotlight paper at ICML 2025 — https://arxiv.org/abs/2412.12119
- Planning with LLMs: Enhanced LLMs with search-based planning techniques to improve multi-step reasoning.
- Asynchronous MCTS: Introduced dynamic virtual counts to balance exploration–exploitation with few simulations.
- Prompt Engineering: Assisted in designing board-game prompts and test-time internal search linearization.
- Technology Stack: Python, Transformer Pre-Training, Supervised Fine-Tuning, Tree-Search Methods
-
Zagreb & Cambridge
Senior Machine Learning Researcher
Cantab Predictive Intelligence (tech startup)
- Key Contribution: Lead a team of four researchers on a few projects running in parallel.
- Behavioral Credit Scoring: Gradient-boosting model for default risk, achieving a market-leading Gini of 75%.
- AI-Driven Marketing: Boosted heart drug sales by 10% via data-driven A/B-tested campaign for pharma client.
- Personalized Newsletter: Built a hybrid recommender (content-based + collaborative); 1.5% CTR in PoC.
- Delivery Delay Estimation: Predicted COVID-era mall delays using ARIMA and supervised learning.
- Technology Stack: Python, PyTorch, PySpark, Databricks, Statsmodels, AWS/Azure, Sklearn, Numpy, Pandas, Git
-
Budapest, Hungary
Machine Learning Researcher
Morgan Stanley
- Key Contribution: Built scalable models for risk, liquidity, and trade execution in financial systems.
- Systemic Risk Model: Built a parallel hill climber heuristic, solving the problem in 3 minutes, averaging 5% from optimal.
- Cash Traceability System: Developed a real-time uncollateralized debt tracker from daily data feeds.
- E-Trading Limits Calibration: Tuned model to block high-risk trades via statistical analysis of client behavior.
- Listed Derivatives Liquidity: Developed a PoC liquidation model driven by intraday futures data.
- Technology Stack: Python, CPLEX, OR-Tools, Q/kdb+, PyQ, SQL, Pandas
-
New York, London & Budapest
Software Engineer
Morgan Stanley
- Annual Grad Program: Participated in a 15-week program for 50 globally selected students.
- Margin Calculator Microservice: Implemented and unit-tested features for NYSE and HGK stock exchanges.
- Technology Stack: Java, C++, Spring Beans, JUnit
-
Zagreb, Croatia
Junior Teaching Assistant
University of Zagreb
- Euclidean Spaces: Delivered problem-solving lectures after achieving the top score in a class of 70.
Education
-
Zurich, Switzerland
PhD in Artificial Intelligence
ETH Zurich
- Key Contribution: Operating a fleet of tens of thousands of agents in real time while satisfying safety constraints.
- Thesis: Safe and Scalable Ride-Sourcing Vehicle Rebalancing: A Constrained Mean-Field RL Approach (defense slides)
- Supervisors: Prof. Francesco Corman and Prof. Andreas Krause
- Research Area: Reinforcement Learning, Multi-Agent Systems, Sequential Decision Making, Data-Driven Algorithms
-
Zagreb, Croatia
MSc in Mathematical Statistics
University of Zagreb
Read Thesis- Thesis: Network Optimization in Railway Transport Planning
- Supervisor: Prof. Marko Vrdoljak
- Distinction: Graduated with honors.
-
Bielefeld, Germany
Visiting Student
University of Bielefeld
- Research Visit: Two semesters funded by Erasmus+ during which I wrote my MSc thesis.
- Host: Prof. Andreas Dress
-
Zagreb, Croatia
BSc in Mathematics
University of Zagreb
Skills
Research
RL
LLMs
MCTS
Safe RL
MFRL
MFC
BO
Programming
Python
C++
SQL
Java
C
Bash & CLI
Q/kdb+
Cloud & Packages
Git
PyTorch
AWS & Databricks
Numpy
Sklearn
Pandas
PySpark
Languages
90%
English
100%
Croatian
10%
German
Challenge
White to move wins.
White to move wins.
Find solutions such that white to move wins.
I am happy to hear your solutions if you can solve it even with the assistance of an engine! I am also open to discussing why many modern engines fail to solve it.
Computer Chess:
I co-developed a Chess Champ Gem for Gemini, which enhances language models with search-based planning techniques to improve multi-step reasoning in board games such as Chess, Chess960, Connect Four, and Hex. It achieves Grandmaster-level performance in Chess with a search move count per decision comparable to human players.
Notable Achievements:
- Silver medalist at the individual Croatian junior (under 20 years) championship in 2011.
- Played in Croatian, German, Hungarian and Swiss leagues.
- A personal best Elo rating of 2585 on www.chess.com ranks me within 3 thousand best players on the platform among over 100 million registered users (99.999% percentile).
- The official Elo rating of 2250 places me among the top 3% of globally registered chess players.