Projects

I have always been curious about starting my own microelectronics project as my academic career has not brought me in contact with low-level computer science so far. At the same time, working from home a lot has made me more aware of air quality. Bringing the two together by building my own air quality monitoring station proved to be a perfect fit.

During my research, I discovered the AirGradient DIY Kit that my solution is based on and where I have been happy contributing to the community ever since. I learned a lot about microelectronics and maker projects in general while building and improving the DIY kit, teaching myself soldering skills along the way. The sensor package is monitoring temperature, humidity, CO₂, and particle matter concentration. Especially being aware of the latter two is very helpful to ensure healthy air conditions at home.

This project was a great opportunity to also dive into the world of home monitoring and automation. To manage the sensor measurements, I set up my own Home Assistant installation and designed a custom ESPHome configuration for the ESP8266 microcontroller. Using Home Assistant, I developed a few home automations that, for example, remind me to vent when necessary or warn me when the Air Quality Index becomes bad. I also integrated some Mi Flora sensors into my Home Assistant setup to monitor plant health and light intensity in my apartment.

Learning web development and creating my own website had been on my bucket list for a long time. Most importantly, I wanted to understand the end-to-end process of website creation, from the concept and design phase throughout implementing a prototype to the final deployment steps.

My goal was to leave out any JavaScript or CSS frameworks that abstract too much from the actual logic. I ended up only using the utility-first framework TailwindCSS and vanilla JS. In a second iteration of the prototype, I integrated the static site generator Hugo to improve maintainability and facilitate future extensions to the site.

The website is hosted on my own virtual private server, which taught me a lot about Linux server administration and web server configuration. Along the way, I also worked with the Python web framework Django, which I did not use for this site, but which I am sure will come in handy for future projects.

Deep reinforcement learning has enabled machine learning models to generalize over the state space of possible observations with sufficient training. However, generalizing over an agent’s action space posed an unsolved research problem at the time of this project. Together with four students from the Neural Information Processing Project course at TU Berlin, we implemented and benchmarked different approaches to master environments with large action spaces. Using the board game Rummikub as our test bed, our ideas focused on reducing the size of the action space and thereby the learning complexity of Rummikub.

Our project goals required us to implement an environment-centric learning setup where the environment controls multiple (potentially different) agents, rather than an agent-centric setup where the agent queries observations from its environment. Since popular RL libraries such as OpenAI Gym did not support environment-centric learning at the time, we designed our own learning environment using the PPO implementation of Tensorforce for our learning agents and the linear solver PuLP to verify player hands and board state.

During my internship in the IT Finance and Controlling department at Bosch North America, I saw an opportunity to improve the department’s financial reports and reduce manual work by automating the reporting process.

Using Microsoft Power BI, including DAX and Power Query, I built a proof-of-concept by redesigning one of the department’s reports and convinced my supervisors to put me in charge of a general revamp of all financial reports. Due to the greater functionality of Power BI compared to the previous spreadsheet-based reports, I decided to also rethink the report contents themselves and worked with internal customers to develop a more insightful and productive experience when using the reports.

The project fit perfectly with the study areas of my Bachelor’s degree, as it included both business and technical aspects. Getting hands-on experience with how data analysis and workflow automation can improve other people’s work motivated me to pursue a Master’s degree with a focus on data science and to work on more technical projects in the future.

My first major coding project was part of the Software Engineering Project course at TU Braunschweig. As a team of six undergraduate students, we developed our own interpretation of the game classic Pac-Man.

Pac-Men has a completely plugin-based system architecture using the Eleanor framework. Leveraging Eleanor allowed us to quickly develop new game content like power-ups or to add features such as a level editor and support for additional input devices. All plugins act independently and can be integrated into the game in an on-demand manner. We followed the traditional waterfall model of software development as part of the course requirements including an official project customer, which taught me a lot about the advantages and disadvantages of different software development models.

At the end of the semester, our team won third place in a student competition among all students participating in the course.