Photo of Martin Krutský

Martin Krutský

PhD student in Explainable AI · Intelligent Data Analysis Lab, FEE, Czech Technical University in Prague

Responsible AI · Mechanistic interpretability · Neurosymbolic AI

I am a PhD student at the Intelligent Data Analysis Lab, Faculty of Electrical Engineering at the Czech Technical University in Prague, supervised by Gustav Šír. I work on making AI systems understandable and accountable: mechanistic interpretability to understand what models actually compute, and responsible AI to make sure explanations serve oversight, regulation, and the people affected by AI decisions. On the technical side, my PhD develops neurosymbolic methods for mechanistic interpretability, such as logic-guided circuit discovery and provably sufficient explanations of graph neural networks. I review for the Neurosymbolic Artificial Intelligence journal and have reviewed for NeurIPS 2026 and ICANN 2026.

I co-founded Responsible AI in Prague, an initiative organizing talks, debates, and the Prague Workshop on Responsible AI. Our research assesses explainability methods for AI governance; we derived XAI desiderata from the EU AI Act. I am a member of the CTU Ethics Committee and have volunteered as an ethics reviewer for NeurIPS.

This autumn, I am an Applied Research Intern at Superhuman (formerly Grammarly), using natural language autoencoders to explain AI-text detectors and AI humanizer models. I took part in the EleutherAI Summer of Open AI Research (SOAR), studying sparse autoencoder pathologies that mimic feature hierarchy. Before and alongside research, I spent 5+ years in industry ML engineering, from NLP models to shipped LLM applications.

News

Focus areas

Responsible AI & governance

Assessing whether explainability methods can support AI oversight and regulation, including XAI desiderata derived from the EU AI Act, plus ethics review and public engagement with the Responsible AI in Prague team.

Mechanistic interpretability

Understanding what neural networks compute, with sparse autoencoders, circuit discovery, and neurosymbolic methods that give provably sufficient explanations of graph neural networks.

Applied ML & LLMs

Five years of industry ML engineering and Python development, from Transformer-based NLP to shipped LLM applications with RAG, tools, and guardrails.

Selected publications

All publications →

Experience

Research

Industry

Education

Full CV (PDF) →