Ulrich Sob
AI Research Scientist at InstaDeep Ltd
Cape Town, Western Cape, South Africa
Summary
👤
Senior🎓
Top SchoolUlrich Sob is an AI research scientist with 11 years of experience bridging physics, signal processing and machine learning, now applying reinforcement learning and LLMs to drug discovery at InstaDeep. He holds a PhD in Astronomy and Astrophysics and built computationally efficient algorithms for radio interferometry, including work on MeerKAT calibration and imaging pipelines. At InstaDeep he has progressed from research engineer to leading AI projects, contributing hands-on to ML systems—evidenced by his backend and value-normalization work in the JAX-based multi-agent RL codebase Mava. Comfortable moving between research and engineering, he combines rigorous statistical thinking with production-focused implementation. Outside work he’s an avid traveller and football player, a balance that keeps his problem-solving grounded and collaborative.
11 years of coding experience
3 years of employment as a software developer
Bachelor’s Degree, Physics and Computer Sciences, Cum Laude, Bachelor’s Degree, Physics and Computer Sciences, Cum Laude at University of BueaDoctor of Philosophy - PhD, Astronomy and Astrophysics, Doctor of Philosophy - PhD, Astronomy and Astrophysics at Rhodes University
English, French
Github Skills (13)
machine-learning10
jax10
python10
reinforcement-learning10
normalization9
data-structure9
normalizing9
algorithm9
data-structures9
normalize9
algorithms9
testing8
pytest8
Programming languages (5)
Github contributions (5)
instadeepai/Mava
Sep 2022 - Jan 2023
🦁 A research-friendly codebase for fast experimentation of multi-agent reinforcement learning in JAX
Role in this project:
Back-end Developer & ML Engineer
Back-end Developer & ML EngineerContributions:97 reviews, 130 commits, 21 PRs in 4 months
Contributions summary:Ulrich implemented target value normalization during training within the IPPO system, modifying the step function in the training module. This involved adding functions to compute running mean, variance, and count, as well as normalization and denormalization functions to process the target values. The changes also integrated a running stats variable to store and update running statistics related to model parameters, which are critical for multi-agent reinforcement learning (MARL) performance. Furthermore, tests were added to validate the value normalization functionality.
multi-agent-systemsmultiagentagentreinforcement-learningreinforcement-learning-agent
ulricharmel/cubiInts
May 2020 - Jan 2022
Radio interferometric calibration solution intervals search package
Contributions:2 releases, 4 PRs, 17 pushes in 1 year 8 months
package-searchcalibrationleafradioresampling
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.