From ML Theory to Efficient AI
I am a researcher in Machine Learning, currently working as an AI Research Engineer at Plumerai , where I work on efficient deep learning for edge devices.
Before my current work, I completed my PhD at the Korteweg-de Vries Institute of the University of Amsterdam, supervised by dr. Tim van Erven. My research there focussed on the mathematical foundations of XAI and Bandit Convex Optimisation. These topics and my contributions are briefly explained below.
For even more detail, you can visit my Research page to find links to my published work. Most code for these papers, and some other projects can be found on my Projects page.
CVResearch interests
Mathematical Foundations of Explainable AIMy main focus of my research was expanding the mathematical understanding of algorithms for eXplainable AI (XAI) and Intepretable Machine Learnin (IML). In this vein we proved that attribution methods cannot have every property you would want it to have at once, that counterfactual methods carries several risks related to the changing of the conditional distribution, and introduced a new method to construct Concept-Based Models using techniques from Causal Representation Learning
- Bandit Optimisation
Together with my collaborators, I was also able to contribute to the field of Bandit Convex Optimisation. We were able to introduce a new algorithm that proved to have the best known Regret bounds in the Stochastic and Adversarial case.
Experience
- 2026 – present
- Plumerai, AI Research Engineer
- 2025
- Booking.com, ML Scientist (3 month research internship)
- 2021–2025
- University of Amsterdam — PhD candidate
- 2018 – 2021
- Amsterdam Data Collective — Data Science Consultant
News
I have started as an AI Research Engineer at Plumerai, working on efficient deep learning for Edge AI.
I successfully defended my PhD in January 2026, which closes my chapter at the University of Amsterdam. [Dissertation (PDF)]
Both our Concept paper and Performative Validity paper will be published at NeurIPS 2025! Camera-ready versions will follow later in October. [Concept paper] [Validity paper]
I will be working at Booking.com the following 3 months as a Machine Learning Scientist. The goal of the project is to put some of the theory we developed for Bandit Optimization problems into practice.
New Preprint online! We introduce a new definition: Performative Validity, which measures the validity of Recourse Explanations after subsequence retrains of the model. We show that only recourse explanations targeting causes of the target labels have this property. [arXiv]
7 older updates
New Preprint online! We construct an estimator that can learn concepts from encodings learned in the Causal Representation Learning framework. [arXiv]
Our new paper 'Online Newton Method for Bandit Convex Optimisation' will be published at COLT 2024. I will also attend, so see you in Edmonton! [arXiv]
I will be present at ICML 2024 to present 'Attribution-based Explanations that Provide Recourse Cannot be Robust' as part of the JMLR-to-conference track!
My second article of my PhD will be published at the conference on AI and statistics (AISTATS)!
My first article has been published in the Journal of Machine Learning Research! [JMLR]
A new preprint is out, where we investigate the possible consequences on the risk by providing recourse! [arXiv]
My first paper, authored together with Rianne de Heide and Tim van Erven, is submitted! [arXiv]