Jan Sosulski

Machine learning scientist with a PhD in computer science focused on robust and sample-efficient models for challenging domains such as noisy speech and neurological signals. Interested in improving ML methods for practical signal processing problems. 🐍🦀

Education

Doctorate PhD in Computer Science (Dr. rer. nat.), University of Freiburg.
In my dissertation Rapidly Adapting Machine Learning Methods for Brain-computer Interfaces I demonstrated how leveraging properties of neurological data increases the sample efficiency of machine learning algorithms for EEG-based brain-computer interfaces (BCIs). This reduced the amount of required training data and even enabled a new unsupervised approach, where no calibration phase is needed anymore for the BCI user. Grade: summa cum laude.
2017 - 2023
Master M.Sc. Computer Science, University of Freiburg.
Thesis: Online Stimulus Parameter Optimization during BCI Experiments, specialization area: Cognitive Technical Systems. Final grade: 1.3.
2014 - 2017
Bachelor B.Sc. Information Systems, Duale Hochschule Baden-Wuerttemberg (DHBW) Mannheim.
Thesis: Exception Handling in Integration Systems based on BPMN, specialization area: Software Engineering. Final grade: 1.6.
2011 - 2014
See jan-sosulski.de/publications for a list of my publications.

Professional experience

Kenbun IT AG AI Engineer / Consultant
I trained, deployed and analyzed robust speech recognition models in various domains. Notably in very noisy domains, where little training data is available and with specialized vocabulary, e.g., medical, maintenance and air traffic communication.
Most used technologies for training and analyses: Python, PyTorch (Lightning), NVIDIA NeMo and for deployment: Rust, ONNX, Kubernetes.
2022 - present
University of Freiburg Research Assistant
Among others: development of novel sample efficient machine learning methods for EEG data and implementation in python / Matlab, supervision of one Bachelor and five Master Theses, conducting EEG experiments in collaboration with the University Medical Center Freiburg.
Most used technologies: MATLAB and Python (esp. sklearn, numpy, mne).
2017 - 2022
SAP SE (Germany and Canada) Dual Student
Implementation of storage planning and middleware algorithms.
Most used technologies: ABAP, JavaScript, Java.
2011 - 2014

Awards

Dissertation For my dissertation I was awarded the "Förderpreis Künstliche Intelligenz" by the "Neue Universitätsstiftung Freiburg". 2024
Best poster At the "8th Graz BCI Conference" I received the best poster award for my paper "Spatial filters for auditory evoked potentials transfer between different experimental conditions". 2019

Skills

Python Mostly used in the context of traditional machine learning (numpy, sklearn) and deep learning (PyTorch). 10 years
Rust In a professional context used for neural network deployment/inference, signal preprocessing and graph algorithms. In a hobby context used for small tools, embedded projects and a bread baking helper program. 6 years
MATLAB Used for EEG signal processing, live BCI/EEG experiments and machine learning. 3 years
R Primarily used to devise exercises for information systems students. 2 years
JavaScript Mostly used as a hobby for servers/daemons that do something with my sports data or as a backend/UI to show some sensor readings. 2 years
C++ Really only used for some image processing exercises and in the context of deep learning with the Caffe framework. 1 year
Other Linux, git, fish, bash, ... and the usual stack of tools. 15 years

Hobbies

Riding some kind of bike Started with mountain bikes, was too scared for actual trails, switched to road biking, was too annoyed by all the cars, settled on gravel biking (for now...).
Running External observers might think I'm on a leisurely stroll, but I really am trying...
Baking bread Yes, I am one of those guys that had to start baking sourdough bread during COVID. In any case, I like multi-step recipes and wrote some planning software so I don't have to get up in the middle of the night just because my dough needs me.
Dancing Mostly ballroom and latin.
Embedded development Don't really care what it is, as long as it somehow measures something or otherwise interacts with the physical world.