I am a Computer Science engineer specializing in hardware-software co-design, machine learning systems, and high-performance algorithmic engineering. Currently pursuing an M.Sc. in Machine Learning at Poznań University of Technology, with a foundation in rigorous competitive programming and production-grade cloud infrastructure.
First semester GPA: 4.9 / 5.0. Conducting master's thesis research in Neurosymbolic AI, focusing on scene decomposition, discrete tokenization, and vector quantization under the supervision of Prof. Krzysztof Krawiec.
Graduated Summa Cum Laude (Bardzo dobry z wyróżnieniem).
Overall GPA: 4.89/5.0. Diploma Exam: 5.0/5.0. Thesis Grade: 5.0/5.0.
Member of GHOST (Group of Horribly Optimistic Statisticians).
GPA: 4.66/5.0. Completed master's-level coursework including Deep Learning, Introduction to Data Science, Machine Learning 1, and Quantum Computing.
Completed rigorous university mathematics coursework beyond the standard engineering curriculum, including Linear Algebra 2 and Mathematical Analysis with Applications 2.
Biological sequence search is fundamentally bottlenecked by DRAM memory bandwidth. For my bachelor's thesis, I designed a hardware acceleration architecture that offloads memory-intensive prefiltering and alignment kernels directly into UPMEM Processing-in-Memory (PIM) DPUs.
This work was accepted and presented at the ISCA 2026 MCCSys workshop. Hardware access for feasibility research and benchmarking was graciously provided by Prof. Onur Mutlu's SAFARI group.
Investigating neurosymbolic scene decomposition and visual reasoning based on the DVP/ASR architecture family. Researching discrete tokenization, vector quantization (VQ), Minimum Description Length (MDL) priors, and differentiable rendering pipelines for scene representation.
Over two years of full-stack and infrastructure engineering, culminating in a complete 8-hour technical handover presentation regarding system architecture upon departure.
Optimized a CLEVR synthetic scene generation pipeline. Profiled and eliminated CPU-side bottlenecks, enabling 100% GPU utilization and reducing rendering time to 5ms per 224x224 image. This pipeline was utilized in a final computer vision study.
Multi-paradigm solvers for the 2-Constraint Knapsack problem. Combined OpenMP CPU fine-grained parallelism with a custom CUDA-accelerated fitness evaluation kernel.
High-performance distributed mutual exclusion implementation using OpenMPI message passing and C++17 thread-safe atomics for local state synchronization.
A collection of moments from hackathons, competitive programming finals, and time spent away from the keyboard.
