Work and Education
Research Experience
Prof. Dr. Vera Demberg's research group
- Saarland University
- August 2020 – February 2021
- Saarbrücken, Germany
Scientific Employee
- Developed a natural language generation model of bar chart descriptions in a few-shot setting.
- Adapted an existing TensorFlow and Python codebase to work with a dataset of annotated bar chart descriptions.
- Automatically measured the correctness and quality of the GPT-2 generated descriptions.
Prof. Dr. Cas Cremers's research group
- CISPA Helmholtz Center for Information Security
- May – July 2019
- Saarbrücken, Germany
Student Assistant
Proposed and analysed multiple strategies for extending the Signal messaging protocol to multiple devices while preserving its security guarantees.
https://cispa.saarland/group/cremers/index.htmlSecure and Privacy-Preserving Systems research group
- Saarland University
- May – August 2016
- Saarbrücken, Germany
Student Assistant
- Theoretical analysis of security protocols for smart homes and IoT.
- Analysis of the WhatsApp encryption protocol.
Teaching Experience
System Security research group
- Saarland University
- November 2017 – May 2018
- Saarbrücken, Germany
Tutor
Organized and taught a weekly tutorial, which included the preparation and grading of minitests. Helped with exam preparation and grading.
https://cispa.saarland/group/rossow/newsProgram of Psycho-Pedagogical Training, Level I
- Department of Teacher Training, West University of Timișoara
- October 2013 – August 2014
- Timișoara, Romania
Student
- Course work about Psycho-Pedagogical Training, how to organise lessons, and hands-on teaching experience.
- Taught high school students introduction to programming.
CoderDojo
- Timisoara Startup Hub
- April – September 2013
- Timișoara, Romania
Mentor
Helped young children learn how to program in Scratch
Education
Saarland University
- Prof. Vreeken
- October 2014 – May 2019
- Saarbrücken, Germany
Master in Computer Science
Thesis: How to be Grim: explaining data at different granularity levels using patterns and structure functions
- Proposed two algorithms, Grim and Brim, which approximate Kolmogorov’s structure function. They find multiple, increasingly detailed explanations for a dataset, that provide a high-level view as well as an in-depth one.
- Improved and extended an existing large codebase in C++.
West University of Timisoara
- Prof. Istrate
- October 2011 – July 2014
- Timișoara, Romania
Bachelor in Computer Science
Thesis: The experimental analysis of several approximation algorithms
- Implemented multiple approximation algorithms in Python for solving the Set Cover and Vertex Cover problems.
- Compared the results with an optimum solution outputted by the mathematical optimization solver Gurobi.
West University of Timisoara
- Prof. Acea
- October 2011 – July 2014
- Timișoara, Romania
Bachelor in Photography
Thesis: People and books
- Course work about Art in general and Photography in particular.
- The final thesis consisted of a series of photographs depicting people reading their chosen books.
Technical Skills
Programming Languages
Advanced: Python, C++, C, LATEX, JavaScript, HTML, CSS, PHP
Intermediate: Bash, R, Java, C#
Packages/libraries
NumPy, Matplotlib, NLTK, Scikit-learn, Gensim, spaCy, Polyglot, Pandas
Machine Learning & NLP
Deep Learning (LSTM, Encoder-Decoder, Transformer) using TensorFlow (Keras)
Named Entity Recognition, Natural Language Generation
Other
Linux, Git, Bootstrap, Docker, MySQL
Languages
Projects
NLG of Bar Chart Descriptions in a Few-Shot Setting
I created an approach for generating bar chart descriptions automatically. Since there are few datasets available for training which contain bar chart descriptions, a secondary goal was to do this with a small training set. Thus, I developed a natural language generation model of bar chart descriptions in a few-shot setting. This included adapting an existing TensorFlow and Python codebase to work with a dataset of annotated bar chart descriptions. I also proposed an automatic measure for computing the correctness and quality of the GPT-2 generated descriptions.
How to be Grim: explaining data at different granularity levels using patterns and structure functions
The focus of my Master's Thesis was finding out whether it is more beneficial to analyse a dataset via multiple models, instead of just a single one. By using Kolmogorov’s structure function we can not only order these on how well they explain the data, but also identify those that are worth looking into. Thus, I created two algorithms, Grim and Brim, for finding these models and through many experiments showed that the algorithms output high-quality, high-level, as well as in-depth explanations.