cv

Machine Learning Research Engineer (Short-term)

Wisent

Built a reproducible LLM steering experimentation pipeline; validated/extended steering implementations; introduced CI and tests; released steering-augmented models.

• Set up HPO and evaluation harness with W&B logging; syntax-level fixes did not translate into reasoning gains.
• Stabilized iteration via CI and unit tests.
• Released two steering-augmented models on Hugging Face; documented evaluation protocol.
• Selected contributions: closed PRs by jfpio — https://github.com/wisent-ai/wisent-guard/pulls?q=is%3Apr+is%3Aclosed+author%3Ajfpio

Machine Learning Researcher

University of Warsaw

Research on RL for multi-agent LLM debate and on predicting narcissism from user-generated text (with Faculty of Psychology).

• Reinforcement Learning for Improving LLM Debate Protocols — building on ‘Training Language Models to Win Debates with Self-Play Improves Judge Accuracy’. Managed LLaMA-family runs on HPC via Singularity; designed synthetic environments and proxy tasks; co-implemented novel RL algorithms and datasets.
• Predicting Narcissism from User-Generated Texts — BERT-based regression and few-shot prompting with SOTA LLMs; transformer models showed promise but data scarcity limited generalization.

Machine Learning Engineer

MIM Solutions

End-to-end ML/NLP work across research and production: dual-encoder contrastive learning, counter-speech generation, vision clustering, disinformation analysis.

• Contrastive Learning for Linking Twitter Posts to News (ERC-funded) — BERT-based dual encoders; experimental design, literature review, development, analysis; collaboration with UW students; preprint: https://arxiv.org/abs/2312.07599
• Automated Counter-Speech to Mitigate Hate Speech — supported student-led project; added engagement metric, expanded literature review and ethics; paper accepted at EMNLP Findings 2024; preprint: https://arxiv.org/abs/2311.16905
• Photo Clustering for advertising client (project value: 25,000 EUR) — embeddings + GroundingDINO + GPT-4; accuracy 99.5% (up from 95%); received positive LinkedIn feedback.
• Disinformation Analysis System for NASK — co-designed ML system; presented at Warsaw Computer Science Days 2024 (rating 4.8/5).
• Public writing on AI & platform governance — Klub Jagielloński article: https://klubjagiellonski.pl/2025/04/23/stanislaw-lem-przewidzial-patologie-x-a-czy-da-sie-uratowac-platforme-elona-muska/
• Ops & infra — trained/evaluated/fine-tuned deep models on GPU clusters; Hydra for configs, Optuna for HPO, Neptune for tracking.

Scout Leader, Board Member

Polish Scouting Association (ZHR) — Mazovian Region

Oversight and QA of summer/winter camps for 2,000+ youth annually; leadership development and community growth.

• Managed camps of 120 participants (ages 11–20), logistics, safety, ~€25k budget; mentored team leaders.
• Led local scout community 4 years; grew from 8 to 10 units, coordinated activities for ~250 members; reduced leader turnover by 50% via new programs and mentorship (~30 leaders).
• Organized public discussions on youth/societal issues with Polish think tanks; video: https://www.youtube.com/watch?v=7XiXUfiU5Ms

Postgraduate (Interrupted)

Sorbonne Université, France

Project Management

M.Sc.

University of Warsaw, Poland

Machine Learning

• Deep Learning
• Natural Language Processing
• Computational Social Choice
• ML in Big Scale
• Explainable ML

B.Sc.

Warsaw University of Technology, Poland

Computer Science

[PL] Stanisław Lem przewidział patologie X-a — a czy da się uratować platformę Elona Muska?

Klub Jagielloński

Public-facing analysis on societal risks of AI-driven content moderation and platform governance.

LLM-generated Responses to Mitigate the Impact of Hate Speech

EMNLP Findings 2024

Automated counter-speech generation against online hate speech targeting Ukrainian refugees. (Authors: Podolak, Łukasik, Balawender, Ossowski, Piotrowski, Bąkowicz, Sankowski)

Linking news to tweets cascades with contrastive learning approach

arXiv preprint

BERT-based dual-encoder methods for linking Twitter posts with news articles around real-world events.