mllm-shap: A Shapley Value Explainability Platform for Text-Audio Multimodal Large Language Models
Muszyński, Pozorski, Ganzha. First author, presented in San Diego.
Engineer | Founder | Researcher
I'm Jakub, an engineer and founder from Warsaw. I like problems where almost right is the same as wrong: detector software at CERN, model inference at TSMC, risk systems at Point72, and two companies of my own. On the research side, I work on explaining what neural networks actually listen to.
An AI energy operator for Polish factories. I lead the edge hardware, the forecasting platform and the battery dispatch behind it.
Production systems across risk, quantitative research and trading, where latency and correctness both matter. Joined as an intern, full-time since January 2026.
With Prof. Maria Ganzha. Digital twins of real industrial sites, with explanations built into every decision the twin makes.
Associated Member of Personnel. Production C++ in the experiment's real-time data-quality system: aging monitoring and ADC-to-MIP calibration for the Fast Interaction Trigger, and its geometry in the event display.
LLM inference with vLLM, tensor parallelism and CPU offloading. A full-stack defect-analysis tool built on fab data.
Clinical notes from speech: streaming recognition plus LLM summarisation. Funded by a university innovation grant, national Enactus champion, top 16 at the World Cup in Bangkok.
Highest honours, GPA 4.70 out of 5. Thesis on explainability for multimodal, multilingual models, graded 5.0.
Muszyński, Pozorski, Ganzha. First author, presented in San Diego.
Muszyński, Walużenicz, Zan, Wrona, Ganzha, Paprzycki, Bădică. First and corresponding author.
Mermer, Muszyński, Możaryn, Rosłon.
Pozorski, Muszyński, Ganzha.
Shapley attributions for models that hear and read at once. Groups audio frames into words before playing the game.
The multi-agent microgrid simulator behind the BDA paper: physical asset models, rolling-horizon planning, agents negotiating a schedule.
A tree-walking interpreter for Lox, compiled to WebAssembly so it runs in your browser.
The same language again, as a bytecode virtual machine in C, with a few extensions the book leaves as exercises.