Model Distillation for Deep Audio Representations
Bachelor thesis investigating how large audio representation models such as BEATs can be distilled into lightweight student networks for deployment on hearable devices.
01 — Work
Work experience
-
Scientific Assistant
Lucerne University of Applied Sciences and Arts
As an AI Research Assistant, I have worked on several applied research projects. I contributed to the automatic transcription and summarization of medical consultations, developing the frontend and implementing LLM-based summary generation.
Another project focused on pedestrian detection and tracking in a public square using drone footage. I also developed a complete information-extraction pipeline for processing photos of pension certificates.
My current work continues my bachelor thesis, where I explore distillation of audio tokenizers for deployment on hearable devices. In addition to research projects, I manage 13 GPU workstations and set up and maintain a small Slurm cluster.
-
Junior Web Developer
V‑ZUG AG
As a Junior Web Developer at V‑ZUG AG, I gained extensive experience in web development using Java, Angular, and Python. Working in a Scrum team, I effectively utilized Git, JIRA, and Confluence. With a strong background in MySQL, Spring MVC, and 3rd Line Support, I contributed to enhancing the company’s web presence.
-
Apprentice Software Developer
V‑ZUG AG
During my apprenticeship, I gained insights into IT and software development. From my second to fourth year, I was involved in improving and troubleshooting the V‑ZUG E‑Commerce platform and developing various data exports. In my final two years, I took on the responsibility of mentoring and training apprentices in web development.
-
Scientific Assistant
Lucerne University of Applied Sciences and Arts
As an AI Research Assistant, I have worked on several applied research projects. I contributed to the automatic transcription and summarization of medical consultations, developing the frontend and implementing LLM-based summary generation.
Another project focused on pedestrian detection and tracking in a public square using drone footage. I also developed a complete information-extraction pipeline for processing photos of pension certificates.
My current work continues my bachelor thesis, where I explore distillation of audio tokenizers for deployment on hearable devices. In addition to research projects, I manage 13 GPU workstations and set up and maintain a small Slurm cluster.
-
Junior Web Developer
V‑ZUG AG
As a Junior Web Developer at V‑ZUG AG, I gained extensive experience in web development using Java, Angular, and Python. Working in a Scrum team, I effectively utilized Git, JIRA, and Confluence. With a strong background in MySQL, Spring MVC, and 3rd Line Support, I contributed to enhancing the company’s web presence.
-
Apprentice Software Developer
V‑ZUG AG
During my apprenticeship, I gained insights into IT and software development. From my second to fourth year, I was involved in improving and troubleshooting the V‑ZUG E‑Commerce platform and developing various data exports. In my final two years, I took on the responsibility of mentoring and training apprentices in web development.
02 — Studies
Education
-
Master of Science in Engineering (Data Science)
Lucerne University of Applied Sciences and Arts (HSLU)
The Master’s profile in Data Science within Information and Communication Technologies at Lucerne University of Applied Sciences and Arts deepens expertise in advanced machine learning and applied research methods. The program focuses on solving complex, real-world problems through project-based work and close collaboration with industry and research partners.
-
Bachelor in Artificial Intelligence and Machine Learning
Lucerne University of Applied Sciences and Arts (HSLU)
The Bachelor in Artificial Intelligence and Machine Learning at Lucerne University of Applied Sciences and Arts provides a strong foundation in machine learning, data science, and AI systems, combining theory with hands-on projects. The program emphasizes applied problem-solving, preparing students to design, train, and deploy AI models in real-world settings.
-
Information Technologist, Application Development
Gewerblich-industrielles Bildungszentrum Zug
Completed an apprenticeship in Application Development at Gewerblich-industrielles Bildungszentrum Zug, earning a Berufsmatura in parallel. Gained broad exposure to software development, covering fundamentals across multiple areas.
03 — Projects
Selected projects
I've worked on a variety of projects, from simple web applications to complex knowledge distillation pipelines. Here are a few of my favorites.
Drone-Based Pedestrian Tracking for City Planning
Industry project analyzing pedestrian behavior in a public square using drone footage and computer vision.
Jass Agent using Monte Carlo Tree Search
A modular Jass-playing agent combining Determinized Monte Carlo Tree Search, neural trump prediction, and rule-based heuristics. The project evaluates algorithmic tradeoffs, MCTS convergence behavior, and distributed search under real-time constraints, culminating in a competitive tournament deployment.
04 — Coursework
I like building things
During my time at university, I completed several project-based modules with concrete deliverables. I like to showcase some of this work here. A few projects are older, and there are many things I would approach differently today, so please don’t be too critical 🙂
-
Deep Learning with CNNs
Deep Learning (DeLearn)
Project for the MSE Deep Learning module, exploring Convolutional Neural Networks for 10-class animal image classification on the iCoSimal V3 dataset. Covers progressive CNN depth experiments, hyperparameter tuning with Optuna (Bayesian search), regularization techniques, and optimizer comparisons, with experiment tracking via Weights & Biases.
-
Bird Call Classification for BirdCLEF 2024
AI Competition
For AICOMP, Flora Gashi, Alvin Veliyath and I participated in the BirdCLEF 2024 challenge, classifying bird sounds from the western ghats into 182 species using deep learning. Focused on audio preprocessing, model training, and evaluation.
-
Jass Bot using Monte Carlo Tree Search
Deep Learning for Games
For DL4G Dan Livingson and I developed a Jass-playing agent using Monte Carlo Tree Search and a regression neural network for trump selection.
-
-
Adventure Architect
Digital Creativity
A ChatGPT-based assistant for Dungeons & Dragons that generates NPCs, spells, monsters, and magical items. Built before structured output was available, with a strong focus on prompt engineering. Due to OpenAI removing unused balance, the application is no longer functional.
-
05 — Web demos
A collection of small interactive tools created using AI. Usually coded in an afternoon.
- Distribution Playground An interactive web app for exploring 24 common probability distributions, from Normal and Beta to Poisson and Negative binomial. You can adjust parameters with sliders, switch between PDF and CDF, and overlay random samples. A built-in Fit game hides a target distribution and challenges you to match it by tuning parameters.
- Regularization Playground An interactive web app for building intuition about regression, overfitting and regularization. You can add and drag data points, choose polynomial, RBF or Fourier features, and compare ridge, lasso and elastic net along the regularization path. A built-in Challenge mode hides a function and asks you to tune the penalty so the model generalizes to unseen test points.
- Reinforcement Learning Playground An interactive web app with sixteen algorithms from Sutton & Barto's Reinforcement Learning: An Introduction, from bandits and dynamic programming to Q-learning, MCTS and actor-critic. Run them live on editable gridworlds, the 10-armed testbed and cart-pole, and watch values, policies and learning curves update as they train.