At Markant AG, I built the retrieval and grounding pipeline for an LLM-based EU regulatory compliance checker, combining dense retrieval, BM25, Reciprocal Rank Fusion, and neural reranking across regulatory documents.
I designed the backend and API to orchestrate evaluation of 122 expert-defined rules, with end-to-end checks running in about 12 seconds per document. I also developed a classifier that distinguishes product from lifestyle images using BiRefNet segmentation and logistic regression.
At RUDEL. GmbH, I developed and trained a multi-label dog emotion recognition model using a ViT backbone and a domain-informed loss function. I integrated it into the app backend and automated training and deployment; the model achieved an F1 score of about 0.64.
At Technische Hochschule Ingolstadt, I developed SmartManPy, a production line simulator for synthetic manufacturing data, RL agents, and quality control. I was a co-first author of a peer-reviewed publication based on this work and presented it at ISM 2024.

