I'm an NLP Researcher at the University of KonstanzResearchI started with a B.Sc. in Business Informatics, added an M.Sc. in Computer Science, and worked as a teaching and research assistant along the way. These days I'm doing my PhD in NLP: teaching machines to notice how Latin texts quote and echo each other, and finding fair ways to evaluate language models. with a history in Software EngineeringEngineeringBefore research I built software as an application developer, mostly with SQL, Python, R and TypeScript. Old habits stuck: whatever I work on tends to end up as a proper package or pipeline. and Data Science ConsultingConsultingAs a data scientist and AI engineer in consulting roles, I help teams turn language models and messy data into things that actually ship., and an interest in large language models and all things machine learning!
Python package
A benchmark and a Python package for detecting intertextual links in Latin literature. The pipeline generates candidates from a source corpus and reranks them with pretrained language models, covering the full range from verbatim quotation to allusion disguised by morphological variation.
Web application
Entity-centric corpus exploration. Detects named entities across a document collection and turns their co-occurrence into a navigable implicit network.
Evaluation Framework
Robust, unified psychometric testing of language models. Runs questionnaires designed for humans against generative models in a reproducible way, so trait claims can be checked rather than taken on faith.
Research code
In-context learning against fine-tuning for topic detection in a low-resource German web setting, on millions of scraped pages collected to measure policy impact.
VSCode Theme
A Solarized theme for VS Code, tuned for contrast that stays comfortable over a long day.
Model Training Code
A fine-tuned RoBERTa checkpoint for multilingual named entity recognition, published on the Hub so cross-lingual entity extraction does not have to start from scratch.
Findings of the Association for Computational Linguistics: EMNLP
A benchmark of 1,490 scholarly attested intertextual links between Latin authors, released with a reference implementation. Distinguishes verbatim quotation from allusion by a deterministic lemma-based rule and compares lexical retrieval, dense retrieval and cross-encoder reranking on the same candidate pool.
Language Resources and Evaluation Conference (LREC)
Questionnaires written for human respondents are increasingly run against generative models to claim traits or benchmark alignment. This framework makes those runs reproducible and comparable, so a reported trait can be re-measured rather than taken on trust.
ACL Student Research Workshop
Political and social scientists study information consumption by classifying millions of visited webpages, which rules out manual labelling. The paper compares fine-tuned encoders against in-context learning in a low-resource German setting, and reports where the cheaper option is good enough.
Companion Proceedings of the Web Conference
A web application for reading a document collection through the people and places in it. Named entities are detected across the corpus and their contextual co-occurrence becomes an implicit network that can be browsed instead of the raw documents.
Preprint
Existing annotation tools assume you are labelling spans of text. Researchers who scrape the web usually need the opposite: a verdict on whether an entire page belongs to a category. Tag-Pag systematises that page-level decision.
EACL Student Research Workshop
A generative model answering a multiple-choice item rarely returns a clean option label. The paper compares classifiers that map free-form responses back onto the intended answer options, a step that quietly decides the outcome of many questionnaire-based evaluations.