Education

PhD in Computer Science · University of Southern Denmark (SDU) and Ordbogen A/S · 2025–present
Department of Mathematics and Computer Science (IMADA), Odense, Denmark
Supervisors: Peter Schneider-Kamp & Lukas Galke Poech
Research focus: distributed machine learning, federated learning, and network-aware training of large-scale neural networks.

MSc in Physics · Ruprecht Karl University of Heidelberg, University of Southern Denmark, and University of Bonn · 2019–2021
Thesis: “Towards Rydberg quantum optics with ultra-cold Yb atoms”
Experimental quantum optics — laser cooling and trapping of ytterbium atoms, Rydberg dressing, and single-photon nonlinear interactions.

Positions

PhD Fellow · University of Southern Denmark · 2025–present
Funded research position at IMADA. Working on gradient synchronisation strategies and communication-efficient training for distributed and federated neural networks.

Researcher · Ordbogen A/S · 2025–present
Applied research on large-scale Danish-language models. Involved in the ordbogen.ai initiative building GDPR-compliant, sovereignty-preserving language AI.

Infrastructure specialist · Ordbogen A/S · 2021–2025
Administrating server and network infrastructure supporting Ordbogen’s dictionary and language-learning platforms. Helped build out the AI training infrastructure that would later become ordbogen.ai, including deployment of the Nordic region’s first NVIDIA HGX B300 server and the surrounding environment.

Publications

FlexMoRE: A Flexible Mixture of Rank-heterogeneous Experts for Efficient Federatedly-trained Large Language Models · 2026
A.B. Pirchert, J. Nielsen, M.H. From, L.G. Poech, P. Schneider-Kamp
arXiv:2602.08818
Low-rank adapters as drop-in expert replacements in MoE architectures; 66 % parameter reduction with improved performance under federated training.

DeToNATION: Decoupled Torch Network-Aware Training on Interlinked Online Nodes · 2025
M.H. From, J. Nielsen, L.G. Poech, P. Schneider-Kamp
Proceedings of the AAAI Conference on Artificial Intelligence (AAAI 2026) · arXiv:2502.06728
Introduced FlexDeMo, a hybrid sharded data-parallel technique that reduces inter-node communication overhead while maintaining validation performance.

A cusp-core-like challenge for modified Newtonian dynamics · 2021
M.H. Eriksen, M.T. Frandsen, M.H. From
Astronomy & Astrophysics 656, A123

PLANCKS 2019 · 2020
H.R. Heebøll, J.M. Benfeldt, M.H. From, et al.
European Journal of Physics 41(3), 034002

Research Interests

  • Distributed machine learning
  • Communication-efficient gradient synchronisation
  • High-performance AI infrastructure
  • Large scale high bandwidth networks

ORCID · Google Scholar · GitHub · LinkedIn