My objective is to make a meaningful contribution to solving the world's environmental, sustainability and social challenges. To achieve this, I aim to combine my deep interest in the environment, human behaviour, and culture with the study of technology, particularly in the area of Artificial Intelligence. I also draw on my passion for art, creativity and serendipity, because I believe the most meaningful innovation happens where science and imagination meet. Feel free to reach out!
digiLab, London, UK
Helping shape the research directions in theLab. Trustworthy AI for high stakes challenges like nuclear fusion, renewable energy and environmental protection, with specialist models that fit on a device.
OceanOS, London, UK
Developed a marine foundation model using geospatial data to support oceanic intelligence and marine conservation.
University of Amsterdam, Amsterdam, NL
Led tutorials for around 30 first- and second-year students, and supported instruction in Programming, Linear Algebra, Calculus and Optimisation, Information Visualization, and study orientation and career.
University of Amsterdam, Amsterdam, NL
Mentored a group of 30 second-year students, providing academic and peer support.
University of Amsterdam, Amsterdam, NL
Graduated cum laude.
University of Cape Town, Cape Town, South Africa
Marine Resources and Geography, Development and Sustainability.
University of Amsterdam, Amsterdam, NL
Specialisation in Artificial Intelligence. Honors programme. GPA 8.02 / 10.
MSc thesis, University of Amsterdam. Supervisor David Wessels, examiner Erik Bekkers. Grade 8.5.
How geometric inductive biases shape weather-model skill and physical fidelity, isolating one design choice at a time on global ERA5. Introduces Platonic-Erwin, a linearly-scaling ball-tree transformer with roto-translation group equivariance. Symmetry helps as an inductive bias but not as a hard constraint: scale buys deterministic skill, geometry buys spectral fidelity, calibration and parameter efficiency.
Published at SIGIR 2025, University of Amsterdam.
Replicated and extended NevIR to assess how IR models handle negation. Only cross-encoders and listwise LLM re-rankers reached moderate performance, still below human level, and fine-tuning across negation datasets generalised poorly.
Deep Learning 2 project, University of Amsterdam.
Proposed a novel class of equivariant transformers, using simplicial complexes and Clifford algebra to capture full symmetry group equivariance.
BSc thesis, University of Amsterdam.
Used GIST features to cluster Amsterdam street scenes, and explored how well those clusters capture perceived attributes of urban environments.
Learning and Deciding project, University of Amsterdam.
Researched CNN-based road defect severity estimation and classification, studied the impact of label noise and proposed robustness strategies.
9 November 2026
The Obsidian vault that runs our coding agents.
With Robin Opdam, on Substack.
24 July 2026
Rules, skills, hooks and an anti-drift system that keep an AI collaborator fast, safe and honest.
24 July 2026
Efficient, trustworthy ML research: guardrail tests, performance callbacks and provenance, wired for Claude Code.
6 May 2025
A guide on scaling machine learning from small to larger training setups.
