In one of my projects with Artifact, we built a data-visualisation chatbot called "hAIdi" for Switzerland Tourism.
It enables non-technical users to gain insights from large amounts of structured data (mainly tourism surveys) by interpreting their requests, generating and executing queries against the data and visualizing the results.

The project was awarded the "Digital Innovation of the Year" award, a category of the Digital Economy Award given out by swissICT.
You can see the hall-of-fame entry here.

For my Master's thesis, I participated in a research project at the ETZ Zurich IVIA Lab under the supervision of Dr. Kenza Amara and Dr. Rita Sevastjanova.

The project aimed at explaining open-ended LLM responses on a concept-level.
To this end, paths in a concept knowledge graph (ConceptNet) are constructed from concepts in the input to concepts in the output, attempting to imitate the model's internal reasoning, i.e., how it made the "leap" from a particular input concept to a particular output concept.

In order to enable evaluation of the obtained explanation paths, I constructed faithfulness metrics based on Shapley values.
I also created a web application that allows (lay) users to construct explanations interactively and explore the underlying KG, built a Python package to work with ConceptNet locally and ran some long-term LLM behaviour analyses.

I own a Snapmaker 2.0. It's basically a modular CNC machine that can do 3D printing, laser cutting/engraving and milling. I bought it for 3D printing, but these days, I mostly use it for milling.

In order to allow myself to use it more like one would a manual mill ("per pedes" = "on foot"), I developed* a web app that can connect to the machine and send gCode commands based on the user's input.
Through this, I can now use the machine even without any CAD/CAM software. There are different modes that emulate features one would find on a manual mill, such as power feed, rotary fixtures or auto-retracting quills for drilling.

You can find the source code on GitHub.

* I had LLMs do most of the coding. Specifications, finetuning, reviews and testing were done by me.

I participated in a research project at the TUM Computer Vision Group for one year during my Bachelor's studies — half a year in the form of a practical course and half a year for my Bachelor's thesis (title: "Deep Reinforcement Learning for Performing Evolutionary Computation").
Both of these projects were under the supervision of Dr. Vladimir Golkov.

The project concerned itself with the optimisation and creation of Evolutionary Algorithms through the use of Deep Reinforcement Learning.
We used modern DRL algorithms such as AWR and SAC to tackle this set of problems.

The Deep Reinforcement Learning algorithm Advantage Weighted Regression (AWR) has been around since October 2019.
The original implementation of the novel algorithm in software was done in Python, using the library TensorFlow.

My own implementation of AWR is written with PyTorch, making it very useful for anyone using this library in their DRL project.
My implementation is the first of its kind on GitHub. It also entails a comprehensive training framework which allows for customisation on multiple levels (hyper-parameters, problems, network architectures) and spares the user the usual overhead of setting all that up.

Below you can see the result of my AWR implementation as applied to the Pendulum problem:

I have written and published a small number of blog posts on Medium, about a range of topics.

One of those posts was called "Getting Started with Deep Reinforcement Learning".
It describes the process one might follow when starting to engage with and tackle projects in DRL. It was marked by the Medium content moderators as a Top Story in Machine Learning and published in Data Driven Investor.

Neuroevolution of Augmenting Topologies (NEAT) is a genetic algorithm that generates artificial neural networks from scratch using principles of evolution.
Some academic implementations have been published, however I wrote this implementation myself, using Python to execute the algorithm and Pyglet for rendering.

This was my first AI project, which is why it's listed here.
Feel free to check it out on GitHub or watch a demo here:

Together with four student colleagues of mine I founded the start-up company "ValveShelf" at the end of 2017 in Passau.

We produced a fully automatic cocktail machine, which could be controlled by a smartphone app which was available via AppStore and PlayStore.
Our business model was renting the machine to private customers and companies hosting small parties and celebrations. People could create and order their own cocktails and have the machine mix it for them.

We also completed the Honours Degree in Digital Technology and Entrepreneurship at the University of Passau with this project.
We eventually terminated our pursuit of the project, favoring our academic careers.

While our website and app don't exist anymore, some pictures and videos survive on our Facebook and Instragram pages.