Sebastian Bodenstein
Research Engineer at DeepMind
London, England, United Kingdom
Summary
👤
Senior🎓
Top SchoolSebastian Bodenstein is a research engineer with a decade of experience applying deep learning to high-impact scientific problems, currently advancing machine learning for nuclear fusion at DeepMind and leading the Tokamax GPU/TPU kernel library. He previously contributed to protein folding projects (AlphaFold 2 and 3) and developed core ML functionality at Wolfram, blending production-quality library work with research-grade modelling. His open-source contributions span foundational frameworks like MXNet and JAX—fixing gradients, attention precision issues, and performance hooks—and implementing AlphaZero within OpenSpiel, reflecting both low-level kernel and algorithmic expertise. Trained as a theoretical physicist (PhD/MSc, University of Cape Town), he brings rigorous mathematical thinking to practical engineering trade-offs and performance-critical systems. An atypical strength is his end-to-end fluency from numerical kernels on GPUs/TPUs to high-level model design, enabling measurable speedups in research tooling and large-scale ML pipelines.
10 years of coding experience
7 years of employment as a software developer
Phd, Theoretical Physics, Phd, Theoretical Physics at University of Cape Town
English, German
Github Skills (29)
algorithm10
algorithms10
python10
attention-mechanism10
mxnet10
machine-learning10
reinforcement-learning10
numpy10
deep-learning10
tensorflow10
jax10
debug9
c-language9
code-optimization9
debugging9
Programming languages (9)
Github contributions (5)
apache/mxnet
May 2016 - Nov 2018
Lightweight, Portable, Flexible Distributed/Mobile Deep Learning with Dynamic, Mutation-aware Dataflow Dep Scheduler; for Python, R, Julia, Scala, Go, Javascript and more
Role in this project:
ML Engineer
ML EngineerContributions:37 commits, 44 PRs, 268 comments in 2 years 6 months
Contributions summary:Sebastian primarily contributed to the development and improvement of machine learning operations within the MXNet deep learning framework. Their work included fixing gradient calculations, refining activation functions like Softmax, and enhancing the Caffe importer. They also added support for sequence layers, improved the instance normalization implementation, and added the ability to set OpenMP threads at runtime and GPU memory querying to the C API. The user's work focused on making the framework more robust and efficient.
pythonschedulerdataflowmutationdata-science
jax-ml/jax
Dec 2020 - Mar 2025
Composable transformations of Python+NumPy programs: differentiate, vectorize, JIT to GPU/TPU, and more
Role in this project:ML Engineer
ML EngineerContributions:55 reviews, 110 comments, 29 issues in 4 years 4 months
Contributions summary:Sebastian made several contributions to the JAX library, primarily focused on improving the functionality and stability of the `jax.nn.dot_product_attention` and `jax.numpy.gradient` functions. This includes fixing bugs, enhancing documentation, and ensuring the consistent use of data types. Furthermore, the user addressed an issue related to inconsistent precision in the forward and backward passes for the attention mechanism. These changes demonstrate a focus on refining core functionalities within the JAX framework.
pytorchpythonjitautomatic-differentiationgpu
Find and Hire Top DevelopersWe’ve analyzed the programming source code of over 60 million software developers on GitHub and scored them by 50,000 skills. Sign-up on Prog,AI to search for software developers.