Curriculum vitae

Bachelor of Computer Science 2023 to 2027

University of Lodz

Researcher, FENG grant Jun 2026 to present

University of Lodz, Faculty of Physics and Applied Informatics, RuFuS lab

“Opracowanie systemu do zintegrowanej analizy widmowej sygnałów mikrofalowych i bioimpedancyjnych w diagnostyce niewydolności serca”, development of a system for integrated spectral analysis of microwave and bioimpedance signals in heart failure diagnostics. 8.29M PLN, TEAM-NET FENG programme of the Foundation for Polish Science; consortium with Warsaw University of Technology led by prof. Teodor Buchner; University of Lodz team led by dr inż. Maciej Ślot.

• Machine learning researcher in an interdisciplinary consortium, extending earlier microwave heart-failure work to joint microwave and bioimpedance signal analysis.
• Also the team’s statistician and simulation analyst: statistical analysis of measurement and clinical data, analysis of electromagnetic simulations, and reducing the compute cost of the simulation pipeline.
• Designed and wrote two ex vivo feasibility studies accepted by the bioethics committee (dielectric characterization of appendices in acute appendicitis and of lymph nodes in lymphadenopathy, pediatric population), run jointly with WUM UCK; responsible for the engineering side of running them.

striga.ai, R&D and machine learning Jun 2026 to present

striga.ai, Warsaw

• Research and development of machine learning methods at an AI source code auditing startup (ISEC spin-off, seed-funded) that finds, validates and reports 0-day vulnerabilities with working proofs of concept.
• First author of a NeurIPS 2026 workshop paper (TAE) showing that how well an LLM appears to detect code vulnerabilities depends heavily on how the benchmark is scored, how the model’s answer is read out, and how the labels were produced. Evaluated 61 open-weight models (1.5B to 36B) and 7 API models on 5,700 vulnerable/fixed code pairs from five public benchmarks and our own data.
• Ran the whole evaluation on the Helios GH200 supercomputer: SLURM jobs, vLLM serving for dozens of models with automatic recovery from out-of-memory and context-length errors, jobs that resume after interruption, answer grading by open LLMs (Llama-3.3-70B, gpt-oss-120b, Qwen3-32B), and linear probes on the hidden activations of 67 models.
• Got 100B+ mixture-of-experts models running across several nodes (vLLM, Ray, NCCL on aarch64); found, documented and patched several vLLM bugs to get there.
• Built a pipeline that turns publicly fixed C/C++ vulnerabilities into training examples with step-by-step reasoning, each confirmed by compiling and crashing the code under ASan/UBSan and checked by a second LLM; built a code-obfuscation tool (tree-sitter, Clang LibTooling) to test whether a detector understands a bug or only recognizes its familiar shape.

Researcher, KPO grant “RuFuS” Mar 2025 to Apr 2026

University of Lodz, Faculty of Physics and Applied Informatics, RuFuS lab

“Implementacja spektroskopii mikrofalowej w diagnostyce niewydolności serca oraz rehabilitacji hipotonii mięśniowej”, implementation of microwave spectroscopy in heart failure diagnostics and rehabilitation of muscular hypotonia. 3.25M PLN, Medical Research Agency call 2024/ABM/03/KPO under the National Recovery Plan; team led by dr inż. Maciej Ślot.

• Sole machine learning researcher in an interdisciplinary team of physicists and medical researchers.
• Developed a PyTorch-based 32-class cardiac condition classifier reaching about 80% accuracy on noisy RF spectroscopy data, above expert human assessment at about 50%.
• Developed and validated a signal reconstruction method demonstrating the feasibility of time-gated analysis from partial magnitude data (about 95% accuracy from 10% of the original data), potentially reducing hardware costs for production deployment.
• Built and maintained a unified analysis toolkit for lab-wide use (6000+ lines refactored into a modular codebase); rewrote performance-critical components in C for real-time visualization of 300+ measurement curves, using matplotlib, scipy, scikit-rf and MatLab.
• Used JAX to implement a non-Gaussian denoising diffusion model from the research literature for experimental data preprocessing, reaching about 85% denoising accuracy.