Niels Rogge is a machine learning engineer on Hugging Face’s Community Science team who makes open research easier to find, reproduce, and use. He integrates research models into Transformers, builds practical educational resources, and develops automated systems that connect published papers with usable models, datasets, and documentation.
Rogge earned a master’s degree in business and information systems engineering at KU Leuven, where 3Blue1Brown’s explanations of neural networks and Andrej Karpathy’s computer-vision lectures drew him into deep learning.
By March 2021, he was an applied AI researcher at Howest in Belgium and an independent Hugging Face contributor. He reimplemented Google’s TAPAS table-question-answering model in PyTorch for the Transformers library and contributed to Microsoft’s LayoutLM for document understanding. His account of those early contributions captures his emphasis on making specialized research architectures usable by other engineers.
He subsequently joined Belgian AI consultancy ML6 while continuing to work part-time with Hugging Face, as he described in a public professional update. His open-source projects include Transformers-Tutorials, runnable examples spanning document understanding, computer vision, and multimodal models, and Vision-Transformer-papers, a curated collection of research extending vision transformers.
Making open research operational
- Research-artifact discoverability: Rogge helps researchers move model weights and datasets from disconnected hosting services onto the Hugging Face Hub, linking them to papers and adding model cards, dataset documentation, task labels, and other searchable metadata.
- Deterministic workflows and autonomous follow-up: His initial system uses scheduled Python workflows and GitHub Actions to inspect papers, find associated repositories, and open relevant issues or documentation improvements. For less predictable researcher follow-up, he combines the Claude Agent SDK, the Hugging Face CLI, open models, and sandboxed execution on Modal. His conference presentation connects this approach to outcomes including PaddleOCR model migration and preparation of a large research dataset for publication.
- Agent evaluation through concrete error analysis: Rogge traces model inputs, outputs, latency, and cost; improves unreliable classifications through majority voting; and reduces invented repository links by supplying agents with the actual GitHub README. His evaluation workflow treats public-facing automation as a system requiring continuous inspection.
