Urs Köster
Software Engineer at Google
San Diego, California, United States
Summary
🤩
Rockstar🎓
Top SchoolUrs Köster is a software engineer and hands-on technical leader with over a decade of industry experience and two decades working with neural networks and probabilistic machine learning. Based in San Diego, he has moved research into production across companies like Google and Cerebras, where he scaled ML teams and led projects from pipeline parallelism to novel training paradigms for drug discovery, CV and NLP. His background spans academia (PhD in Computer Science) to leadership roles at Intel, where he drove limited-precision formats and high-performance deep learning in the Nervana neon framework. A pragmatic engineer, Urs contributes to major open-source projects such as TensorFlow Probability and Nervana/neon, optimizing samplers and GPU kernels for substantial performance gains. Colleagues know him for blending rigorous research instincts with production-first engineering, often surfacing numerical-stability and precision tricks that aren’t obvious from model code alone.
12 years of coding experience
10 years of employment as a software developer
Master in Science, Physics, Honors of the first Class, Master in Science, Physics, Honors of the first Class at University of St. AndrewsPhD, Computer Science, Excimia cum Laude, PhD, Computer Science, Excimia cum Laude at University of Helsinki
English, German
Github Skills (15)
cuda10
statistics10
gpu-programming10
machine-learning10
probabilistic-programming10
deep-learning10
tensorflow10
batch-normalization10
python10
optimization9
back-end-development9
bayesian-methods9
deeplearning-ai8
numpy8
data-science7
Programming languages (3)
Github contributions (5)
NervanaSystems/neon
May 2015 - May 2017
Intel® Nervana™ reference deep learning framework committed to best performance on all hardware
Role in this project:
Back-end Developer & ML Engineer
Back-end Developer & ML EngineerContributions:57 commits, 2 PRs, 2 pushes in 2 years
Contributions summary:Urs implemented GPU backend calls for fprop and bprop passes of Batch Normalization, leading to improved performance. The user also made changes to the `neon/backends/gpu.py` file that involved the application of Adam optimization, which provides an updated rule for AdaDelta. They also modified the imageset.py file by adding minibatch producer and del_mini_batch_producer functions. Finally, they incorporated batch normalization forward and backward passes, indicating a focus on deep learning framework optimization.
deep-learningbest-performanceintelmachine-learningperformance
tensorflow/probability
Sep 2020 - Dec 2022
Probabilistic reasoning and statistical analysis in TensorFlow
Role in this project:ML Engineer
ML EngineerContributions:17 commits, 4 comments, 3 issues in 2 years 3 months
Contributions summary:Urs primarily contributed to improving the performance and accuracy of probabilistic models within the TensorFlow Probability library. They optimized the Gamma sampler, resulting in significant speed improvements. Additionally, the user made internal changes, cleaned up code, and updated Numpy linear operators. They also removed and adjusted distributions in tests related to numerical stability.
statisticspythonprobabilistic-reasoningdata-sciencedeep-learning
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.