I'm doing an MSc in Data Science at Warsaw University of Technology, where I also finished my BSc with honors. I write Python and SQL for a living, and in my free time I try to work out why large models say what they say.

My bachelor's thesis turned into a paper on explaining text-audio language models with Shapley values. Along the way I got into the less glamorous part of research: packaging code so other people can install it, rerun it and get the same numbers. I'm also co-founder of Botuj, an AI chat and voice assistant for customer support.

My work so far is about explaining multimodal language models. Newest first.

mllm-shap: A Shapley Value Explainability Platform for Text-Audio Multimodal Large Language Models Jakub Muszyński, Paweł Pozorski, Maria Ganzha ACL 2026, System Demonstrations, pp. 387–396

An open-source platform that extends Shapley-value attribution from text-only LLMs to models that take both text and audio. It masks each modality separately, tracks multi-turn conversations, and groups audio tokens by phonetic alignment, which shrinks the number of coalitions to evaluate by 10–50×.

Paper · PDF · Code · Docs · PyPI

@inproceedings{muszynski-etal-2026-mllm,
    title = "mllm-shap: A Shapley Value Explainability Platform for Text-Audio Multimodal Large Language Models",
    author = "Muszy{\'n}ski, Jakub and Pozorski, Pawe{\l} and Ganzha, Maria",
    editor = "Durrett, Greg and Jian, Ping",
    booktitle = "Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations)",
    month = jul,
    year = "2026",
    address = "San Diego, California",
    publisher = "Association for Computational Linguistics",
    pages = "387--396",
    doi = "10.18653/v1/2026.acl-demo.38"
}

Full list on ORCID.

• MLLM-SHAP is the library behind the paper. It's now growing into a general tool for conversational models: Shapley and Banzhaf attributions, reproducible run archives, and backends for Hugging Face, LM Studio and OpenAI-compatible APIs.
• mllm-xai is the research workspace around it, with experiments and write-ups driven by YAML configs.
• happy-housing is a side project for the Warsaw housing market. It estimates fair prices and shows explainable maps built only from open data.

• AutoPrep (PyPI): automated preprocessing for ML, with explainability and reports.
• lm-studio-dash: a dashboard for model usage and hardware load when running LM Studio on a Mac.
• python-package-template and python-project-template: the uv-based scaffolds I start new packages and experiments from.
• University work: speech recognition with transformers, anomaly detection, clustering, CINIC-10 classification and a few more, all in my repositories.

How I set up projects. Almost everything is Python 3.12+ managed with uv, linted with ruff and mypy, tested with pytest, checked by pre-commit hooks and built in GitHub Actions, with Sphinx for docs and a Makefile for the common commands. Experiments are tracked in MLflow and configured from YAML. For tabular work I reach for LightGBM, XGBoost and SHAP; for geo and data-heavy side projects, DuckDB, GeoPandas and OSMnx. Cursor, Claude Code and Codex all have a place in the repos, with AGENTS.md files and a notes vault so they share context.