I'm a PhD candidate at TU Dortmund / Lamarr Institute, working with JProf. Dr. Matthias Feurer and Prof. Dr. Katharina Eggensperger on making LLM agents smarter, without touching the model weights.

Outside research, I'm a member of the Young AI Leaders Dortmund Chapter.

• Automated prompt optimisation for LLM agents: building frameworks that decompose agentic pipelines and iteratively refine prompts using environment feedback. Think evolutionary search, but driven by what the agent actually did wrong.
• Automated agentic topology search: figuring out how to automatically induce the right agent architecture for a given task, rather than hand-designing it every time.

At the ML Lab in Freiburg (2024–2026), I built agentic LLM pipelines for autonomous tabular data augmentation and explored multi-armed bandit methods for feature engineering decisions. Before that I spent time building data distillation and benchmarking pipelines for LLM training on legal text, which eventually turned into a paper.

Environment-Grounded Automated Prompt Optimization for LLM Game Agents (code) An automated prompt optimisation framework for LLM agents that decomposes the observation-to-action pipeline into a descriptor and action agent, and iteratively refines each module's prompts via an LLM-driven evolutionary loop guided by environment returns. Evaluated on the BALROG benchmark: on PutNext, a multi-step coordination task where the baseline scores 0%, the framework reaches 72.5% using the same underlying model.

A Llama Walks into the Bar: LLM Fine-Tuning for Legal Reasoning (code) Fine-tuned Llama 2 7B (Q-LoRA) on US Bar Exam questions using automated data distillation via Llama 3 70B, achieving 3–4x performance gains over the base model. Includes full training and evaluation pipelines on HPC infrastructure.

Smart SLURM Job Handler A Python + SQLite tool for automated job management on HPC SLURM clusters. Handles job queuing, retries, and status tracking without manual intervention. Built out of practical necessity while running large-scale experiments.

Model Performance Analysis Dashboard (live) Interactive Streamlit app for visualising model performance metrics and learning curves. Built as a companion tool for the bar-llama paper.

MCQ Bias Analyzer Streamlit dashboard for analysing answer selection biases in multiple-choice tasks. Compares predicted vs. ground truth label distributions and visualises confusion patterns. Used to analyse fine-tuning effects in the bar exam paper.

Website · Google Scholar · LinkedIn · Email