Associate professor at FIT CTU in Prague · Co-founder of Recombee
I work on machine learning, and much of it ends up in production at Recombee. I try to use AI to make things around us a little better: recommenders that help people find what they would otherwise miss, AI that is safe for children at school, climate models for cities and forests, and AI agents that work for people.
Things I build with students and colleagues. Most of them are open and online.
With Jakub Matušek we measure how a shopper's agent and a platform's agent should divide the work, and how easily they are manipulated (IJCCI 2026). With Jan Fikeis we show that general-purpose moderation misses most harms aimed at children (WOAH @ EMNLP 2026).
Twenty years of ACM RecSys research on one map: 2,212 papers grouped into 34 regions by what they are about, measured trends, and open problems drafted from the literature.
A book about recommender systems that personalises itself with one and lets readers extend it. Readers choose how content is served and told, and watch the mechanisms act on their own behaviour. Demo at RecSys 2026, with Eva Nečasová.
The interactive knowledge base of our Horizon Europe project: nine pilot regions and their action plans, EU rural policy, open geodata and the JackDaw GeoAI agent. Twelve chapters, reading missions and mini-games, every claim traced to a public deliverable, and an MCP server so that AI agents can read it too.
A slide editor in one HTML file. The deck stays a folder of markdown and SVG figures, legible to you, to git and to the AI model you edit it with. The research slides further down this page are made in it.
An AI assistant for classrooms by AI dětem. I advise on the technology and on keeping it safe for children; lesson decks for teachers are built in SlAIdy.
Interactive p-books for children aged 8–15, made with AI dětem: how recommendations work (in English and Czech), how the digital world works, and how the internet works.
Research code and data released with our papers.
Things made for teaching and for founders.
Companies are how research reaches people. I co-founded three of them and spent seven years as the faculty's vice-dean for industry cooperation.
AI recommendation and search as a service, founded in 2015 with PhD students from my group. Good recommendations help people discover content they would otherwise miss, and help companies make their online products better. I am responsible for the research direction and the link to the university. Together we run Recombee Lab, a joint research lab with the faculty, and Recombee funds industry PhD scholarships; methods such as ELSA, beeFormer and sparse embedding compression came out of this collaboration. The company is led by Tomáš Řehořek, my former PhD student.
500+ customers in 40+ countries and 10,000+ sites and apps, from media such as The Telegraph and DAZN to e-commerce and marketplaces. Annual recurring revenue of about USD 8M.
Joint research that Recombee released as open source:
Practical experience during studies matters more every year, and it is hard to get. In EduMatch every student has an AI career assistant that finds internships and jobs that fit, does the pre-screening for them and spares them painful interviews they had no chance in. A spin-off of experts.ai.
A search engine for university expertise. Research teams show their results, licences and services, and faculties embed the search on their websites so that companies find the right people and equipment.
What I work on with PhD students and colleagues at FIT CTU and with the research team at Recombee: recommendation and search, trustworthy AI agents, machine learning for weather and climate, and AI for science and learning.
beeFormer puts a language model inside the training loop of a recommender, and recall on books it has never seen more than doubles (RecSys 2024). CompresSAE makes embeddings 12× smaller and runs in production (RecSys 2025); Sparse ELSA makes them readable (The Web Conference 2026); at The Telegraph, editors use the same idea to see which reader groups engage with which topics.
WeatherFusionNet predicts rain eight hours ahead from satellite images and won Weather4cast at NeurIPS 2022. In the Horizon Europe project FOCAL we bring climate projections from 12.5 km down to under 500 m, with error bars, for city heat and forest growth. This line of work led group members to found Tempest Labs in 2025.
The JackDaw GeoAI agent and AgriBot answer farmers' questions with local data. Our guards catch advice that breaks a local rule (44 → 86 % of harmful replies, GeoAI 2026) and harms aimed at children (38 → 99 % recall on our benchmark, WOAH 2026). And when both the buyer and the seller are agents, we measure who should do the work (IJCCI 2026).
In AutoResearch, agents run the experiment loop and a hidden test set confirms every result. Recommender maths chooses fluid-dynamics models with HZDR Dresden, and machine learning tunes laser–plasma experiments with ELI Beamlines and UC Berkeley. ResearchCoach, our AI mentor, comments on students' research drafts within minutes.
With Mária Bieliková and Markus Schedl I was general co-chair of the main conference of the field: more than a thousand attendees, about sixty percent of them from industry, and 440 research submissions. Rodrigo Alves from our group chaired the industry track.
Program committees. Senior PC member of ACM RecSys 2024 and 2026 and The Web Conference 2026; PC member of RecSys, KDD, ICANN, IJCNN and other conferences.
Journal reviews. Neural Networks, Information Sciences, IEEE Transactions on Neural Networks and Learning Systems, Machine Learning, ACM Transactions on Knowledge Discovery from Data, Neural Computing and Applications.
Roles. ACM RecSys Steering Committee; General Co-Chair of RecSys 2025 in Prague; Council of the Institute of Computer Science of the Czech Academy of Sciences. Doctoral examiner and thesis reviewer at universities in Prague, Brno, Ostrava and Bratislava.
European projects where I am the Principal Investigator for CTU.
Place-based development of rural and peri-urban regions, with 20 partners from 13 countries, 9 pilot regions and 26 open-call teams. Our group builds the AI layer: the JackDaw GeoAI agent, AgriBot, which tells farmers which subsidies apply to their parcels, and safety guards that know the local rules.
Local climate information for adaptation planning. Our group leads the AI models: super-resolution networks and Bayesian neural fields bring climate projections from 12.5 km to under 500 m and say how sure they are. They feed two pilots, urban heat in Constance and tree growth in Czech forests.
Interreg Danube, January 2024 – June 2026. Technology transfer, company building and ICT innovation across the Danube region: CTU was the lead partner of 16 partners from 9 countries, working with digital innovation hubs and experts.ai. In January 2026 we hosted the ICT Technology Transfer Days at FIT CTU: more than 25 incubated startups from across the Danube region pitched to investors in three tracks. Project page
Two organisations I co-founded: one brings AI education to children, the other builds the AI ecosystem in Prague.
“AI for children”: a non-profit that introduces children and their teachers to artificial intelligence in an engaging and responsible way. It is an important area that few people work on, and we want to take it beyond Czechia. If that makes sense to you and you would like to help, get in touch.
A non-profit initiative of the City of Prague, Charles University, the Czech Academy of Sciences and CTU. For years we have been building the local AI ecosystem and preparing our society for a big change, the arrival of AI superintelligence, so that the region benefits from it and turns early adoption into future wealth.
Academia, companies and non-profits, side by side.
Advanced data mining with a focus on modern recommender systems and search, and the theory behind them, such as the bias–variance decomposition and modelling a model's confidence.
Modern AI architectures and optimisation methods, from the theory to working implementations. The course has its own living p-book.
Discussions of current AI research and breakthrough papers, where students practise reading critically and presenting research.
I introduced the master's courses Data Preprocessing, Data Mining Algorithms and Computational Intelligence Methods and the bachelor's course Knowledge Discovery from Data at FIT CTU and the Data Mining course at FEE CTU.
AI in my teaching. I try out new ways of teaching together with the students, in courses with dozens of students each semester. NI-MVI and the machine learning seminar run on a collaborative course platform where students extend the course text, and ResearchCoach, an AI mentor we built, comments on their research drafts within minutes.
Want to work with us? Thesis and PhD topics · get in touch
Selected papers, newest first. For a complete list see Google Scholar and DBLP.
If something on this page is close to what you do, get in touch. A few lines are enough; we can sort out the details later.
Doctoral topics follow my research lines: recommendation and search, trustworthy AI agents and language models, AI for climate and science, and AI for learning. Experience from a company is an advantage.
Joint papers, Horizon Europe and national proposals, and contract research where we have shown results: recommendation and search, language-model agents and their safety, machine learning for weather and climate, and AI in education.
Topics I know best: recommender systems and search in practice, AI agents and their safety (also for children), machine learning for weather and climate, AI in education, and how university research turns into companies.
Researchers, developers, designers and teachers are welcome in Tiny (AI for schools), p-books, SlAIdy, AutoResearch, colAIborate and EduMatch, whether the project is academic, non-profit or commercial. Tell me what you would like to do.
Teaching children about AI is important and still neglected. AI dětem works with Czech schools, and we want to scale it globally. Schools, funders, translators and partners abroad: let's talk.
I supervise a limited number of theses, for independent students. A topic of your own is welcome if it builds on our research; it should have research potential and a result you can measure.
Name, e-mail and a few lines about what you have in mind are enough. I will reply and we can sort out the rest then.
