I’m a Dane and a PhD fellow at DTU (DDSA PhD Fellow), currently visiting Stanford’s STAIR Lab. I work at the intersection of trustworthy ML, causality, and neuroscience/healthcare—with a slightly more theory / math tilt than average, but always grounded in real data.

• Visiting researcher at Stanford’s STAIR Lab focusing on trustworthy AI.
• Interested in AI alignment and what’s happening inside neural networks—how representations form, align, and sometimes fail in ways that matter for robustness and interpretability. This also motivates my recent work on spectral PLS (including missing-data–induced phase transitions) and on mental rotation as a probe.
• Developing PatternLocal from our NeurIPS paper “Minimizing False-Positive Attributions in Explanations of Non-Linear Models”—aimed at reducing spurious attributions while keeping local explanations faithful.
• Working on the new Danish AI supercomputer Gefion to use Neural Architecture Search (NAS) to find the optimal EEG foundation model.

• PhD candidate at DTU (Applied Mathematics & Computer Science) working on the Causal Approach to Trustworthy AI in Healthcare project (DDSA page).
• Stanford STAIR Lab (visiting), with a focus on trustworthy AI and applications to neuroimaging.
• DDSA fellow and Young Academy Panel member representing early-career data scientists.
• Pioneer Centre for AI collaboratories on signals, decoding, and causal explainability.
• Exchange at ETH Zürich, where I studied mathematics and collaborated with the University of Zurich—work that led to SPEED (DTU Orbit publication).
• Worked with BrainCapture, a startup bringing accessible EEG solutions worldwide.
• Co-founded Copenhagen MedTech, which (among many events) ran a Google Cloud Hackathon.
• Competitive programming enthusiast (ICPC style contests).

• Minimizing False-Positive Attributions in Explanations of Non-Linear Models (NeurIPS 2025) — introduces PatternLocal to suppress false-positive feature attributions.
• SPEED: Scalable Preprocessing of EEG Data for Self-Supervised Learning (MLSP 2024) — a pipeline that stabilizes self-supervised training on EEG and improves downstream performance.
• Concept-Based Explainability for an EEG Transformer Model (MLSP 2023) — adapts concept activation vectors to EEG transformers.
• Large Vision Models Can Solve Mental Rotation Problems (ICASSP 2026) — shows self-supervised ViTs capture geometric structure better than supervised models, and that intermediate layers outperform final layers.
• Missing-Data-Induced Phase Transitions in Spectral PLS for Multimodal Learning — forthcoming preprint (links coming soon).