Derck Prinzhorn

I am an MSc Artificial Intelligence thesis student at the University of Amsterdam (UvA). Alongside my studies, I co-founded Wisr, an education technology startup, and work as a freelance AI researcher .

My interests are broad, but I tend to focus on high-risk AI applications. This includes topics such as responsible AI, AI safety, conformal prediction, medical AI, and AI for education.

I worked as AI Architect at the Dutch Police, where I designed AI and MLOps reference architectures with a focus on security and governance. In addition, I completed several research internships across academia, healthcare, and applied AI, including projects at the Netherlands Cancer Institute, Deltares, the University of Amsterdam, and the Supervised Program for Alignment Research.

I began my Bachelor's in Artificial Intelligence at the UvA in 2020, graduating cum laude. During this time, I was active as a Teaching Assistant, worked as a software engineer at LeerLevels, and joined the Dutch Nao Team robotics group. My bachelor thesis on uncertainty quantification, in which I benchmarked adaptive conformal prediction methods for time series, was awarded the Amsterdam AI Thesis Award.

Email  |  GitHub  |  Google Scholar  |  LinkedIn  |  CV

Highlights / News

Research

I'm interested in AI safety, conformal prediction, medical AI and edtech.

HIVE: A Hyperbolic Interactive Visualization Explorer for Representation Learning

We present HIVE, an interactive dashboard for exploring and interpreting hyperbolic embeddings in deep learning. HIVE offers Poincaré projections, custom dimensionality reduction, and four interaction modes (comparison, traversal, tree, neighbors). A user study shows that HIVE supports practical analysis of hierarchical structure, with potential applications in reinforcement learning and graph discovery.

Morpheus: Benchmarking Physical Reasoning of Video Generative Models with Real Physical Experiments

Recent advances in video generation raise hopes that these models possess world modeling capabilities. We introduce Morpheus, a benchmark of 80 real-world videos for evaluating whether generative models respect physical conservation laws. Using physics-informed metrics, we find that while modern models generate visually appealing videos, they struggle to encode fundamental physical principles.

Conformal time series decomposition with component-wise exchangeability

We present a novel use of conformal prediction for time series forecasting that incorporates time series decomposition.

Reproducibility study of FairAC

This work reproduces Guo et al. (2023) on “Fair Attribute Completion on Graph with Missing Attributes,” confirming its claims, generalizability, and fairness improvements, with a refactored codebase.

Design and source code from Jon Barron's website