👋 Hey, it's me
We investigate the zero-shot, few-shot and fine-tuning capabilities of Large Language Models (LLMs) to classify German texts according to the Common European Framework of Reference for Languages (CEFR). Published at the IEEE FLLM 2025 conference.
View on arXivA cooperative multi-agent reinforcement learning framework in which every electric truck decides where and how much to charge on the real German highway network. Trained with MAPPO and attention over a graph-structured observation, in a custom simulation with real charging infrastructure, EU driver rest rules, and station failures. At 5,000 trucks, the learned policy cuts mean total travel time by 17.9% and fleet-wide queue time by 85.5% compared to a greedy A* baseline, without any additional chargers. It also generalizes from 100 to 35,000 trucks without retraining.
Simulation of electric trucks and charging infrastructure in a logistics network using realistic battery and charging models, driver behavior, and route planning algorithms. Developed in Python with a focus on modularity and extensibility. Project @ Autonomous Intelligent Systems Group.
Evaluates zero-shot, few-shot, and fine-tuned LLMs on German texts across all CEFR levels (A1–C2). Includes a custom evaluation framework and benchmark for comparative analysis, and laid the groundwork for the IEEE FLLM 2025 publication.
View ThesisA comprehensive dataset for German language proficiency classification, containing about 1,500 texts across all CEFR levels (A1-C2). Compiled from multiple sources including the FALKO and MERLIN corpora.
View GitHubA React-based web application to help students at the University of Münster select courses from the "Kerninformatik" area. Features include a visual schedule and automatic conflict detection.
Open Tool