Research Focus My research interests lie at the intersection of AI-based Planning, decision-making in complex environments, and Reinforcement Learning. I am driven by the challenge of advancing the reasoning capabilities of AI algorithms and exploring the strategic thinking that emerges within Game Intelligence. To further pursue this direction, I have recently begun delving in Meta-RL, and Continual Learning to develop algorithms capable of solving problem instances under, broadly speaking, heavy domain shifts across different axes.
Professional Experience
- Senior AI Consultant IV, Ernst & Young2025 - presentWarsaw, Poland
- Deep Learning Researcher, DeepFlare2022 - 2025Warsaw, Poland
- Data Scientist Intern, MicrosoftJul 2022 - Oct 2022Dublin, Ireland
- Teaching Assistant, Uni of Warsaw2021 - 2022Warsaw, Poland
- Deep Learning Intern, NvidiaJul 2021 - Oct 2021Warsaw, Poland
- Software Engineering Intern, MicrosoftApr 2021 - Jun 2021Dublin, Ireland
- Deep Learning Intern, NvidiaJun 2020 - Sep 2020Warsaw, Poland
- Software Developer Intern AI, SamsungJul 2019 - Sep 2019Warsaw, Poland
Education
- M.Sc. Machine Learning, University of WarsawWarsaw, Poland 2021 - 2023top 5% students · graduated with honorsKey activities:
- ICLR A* oral top-5%, onsite presentation on RL and planning
- co-leader ML in PL Conference 2021
- B.Sc. Computer Science, University of WarsawWarsaw, Poland 2018 - 2021top 5% studentsKey activities:
- Published paper on efficient transformers with Google Research.
- President, Machine Learning Society at UW.
Computer Science Background My experience spans the entire AI stack: from low-level C++/CUDA engineering (NVIDIA Triton Inference Server) and scalable distributed systems (Microsoft Omex), through MLOps and ML library development (CaRL: Deep RL library calibrated for planning and search), to advanced applied & research innovation across industry and academia. I have worked on deep learning for vaccine discovery (DeepFlare), 3D computer vision algorithms (Samsung), and built my own fastest model parallelism algorithm at the time for extremely large NLP models (NVIDIA). I was the first Microsoft Ireland intern to have a paper accepted at MLADS on explainable AI. My bachelor’s thesis with Google Research set state-of-the-art benchmarks for efficient transformers in long-sequence prediction. During my master’s, I developed Adaptive Subgoal Search (AdaSubS), a novel search algorithm for efficient reinforcement learning under low computational budgets-presented onsite as an ICLR 2023 Top-5% Oral (first such achievement from Poland). Most recently, my research on hierarchical search landscapes was awarded Best Poster at EEML 2025.
Community I am an active member of the AI community, particularly within the ML in PL Association, a non-profit advancing the machine learning community in Poland and across Central & Eastern Europe. Since 2020, I have served as Scientific Program Officer across six annual editions-curating a high-impact invited speaker lineup-and had the honor to be co-Leader of the ML in PL 2021 Conference. In my free time, I enjoy mountain hiking and motorization.
You can view my full CV here.
Selected Publications
CoRL 2025 workshop
Paper Code Website Colab
Benchmark and toolkit for evaluating VLMs' sense of progress in robotics via Value-Order Correlation (VOC); enables automated dataset curation from videos.
Cite
ICLR 2024 (Generative Models for Decision Making)
Paper Code
Empirical analysis of properties shaping hierarchical search in combinatorial reasoning; guidelines for robust comparison + future design.
Cite
ICLR 2023 (Top-5%, Oral)
Paper Code Website Colab
AdaSubS adaptively adjusts planning horizon via diverse subgoals + fast reachability filtering; efficient on Sokoban, Rubik's Cube, INT.
Cite
NAACL 2022
Paper
Hourglass: a hierarchical Transformer with down/upsampling layers that improves long-sequence modeling efficiency; SOTA on ImageNet32 and strong enwik8 performance.
Cite
Other Papers
Microsoft MLADS 2022
Had the privilege to lead explainable AI research at Microsoft Ireland during an internship, utilizing petabytes of data to develop interpretable machine learning solutions for both theoretical and practical applications, presented at the internal MLADS conference.
Cite
ICLR 2023, MLDD Workshop
Paper
This study pinpoints evaluation pitfalls in MHC-I presentation prediction and proposes a unified framework to standardize methodology. It also introduces a transformer model trained on interspecies data, markedly improving peptide–MHC-I binding accuracy and generalization across diverse peptides, alleles, and proteins.
Cite
Selected Projects
Co-author & contributor
OpenGVL is an open benchmark and toolkit for measuring how well vision–language models understand temporal task progress in robotics, enabling automatic dataset curation by predicting per-frame completion from videos.
- Problem: Equip robots with a sense of progress for better learning and decision-making.
- Metric (VOC): Value‑Order Correlation - Spearman rank correlation between predicted progress ordering and true time order.
- Few‑shot prompting: Uses ordered context episodes to guide predictions on shuffled frames.
- Contamination control: Hidden tasks with curated demos; 100% completion rate datasets used for evaluation.
- Unified interface: Standardized prompts, data loaders, and configs (Hydra) across open and closed VLMs.
Lead author & maintainer
CaRL is an open-source library for scalable offline and online reinforcement/imitation learning in combinatorial planning problems.
- Supports environments like Sokoban, NPuzzle, Rubik, and INT.
- Includes 35+ open-source models (Generator, Value, Policy, CLLP).
- Enables distributed experiments on SLURM clusters and local machines.
- Interactive Jupyter notebooks for research and reproducibility.
- Used in multiple peer-reviewed papers.