Ben Burtenshaw is a machine-learning engineer at Hugging Face and coauthor of the Hugging Face Agents Course. He builds open tools and educational resources that extend coding agents into GPU optimization, model training, collaborative research, and software maintenance.
Earlier, his doctoral research at the University of Antwerp examined how natural-language-processing systems could identify toxic language in conversations involving adolescents. His research on interactive storytelling for children explored conversational technology for collaborative narrative and learning, while subsequent work on open human feedback investigated how open-source communities could share the human feedback essential to developing and improving AI systems.
At Hugging Face, he helped develop educational tools for working with open models and became one of four credited authors of the Hugging Face Agents Course, which teaches agent frameworks, observability, evaluation, and practical implementation.
- Agent-written CUDA kernels. Burtenshaw coauthored a practical approach to agent-generated GPU kernels that combines executable examples, hardware-specific benchmarks, integration tests, and distribution through the Hugging Face Hub. His emphasis on GPU memory bottlenecks explains why optimized kernels can accelerate workloads by performing more computation for each costly transfer between memory and processor.
- Inspectable agent skills. He treats skills as version-controlled files containing task instructions, examples, and executable checks. Project maintainers can update these materials alongside their software, while comparative evaluations reveal when cheaper or open models can handle specialized engineering work.
- Autonomous open-source research labs. His multiautoresearch project divides model-improvement experiments among agents responsible for literature research, planning, implementation, review, and reporting. Git branches preserve experimental changes, while Trackio dashboards and accessible Parquet data make results inspectable by both engineers and software.
Burtenshaw continues developing resources for training agents and educational material on agent evaluation, pursuing AI infrastructure whose experiments, data, and engineering decisions remain open to inspection.
