Curriculum Vitae
Machine Learning Engineer at Ergon Informatik AG in Zurich. I build and evaluate language-model systems and take them to production, with a research background in efficient multilingual NLP from KIT and EPFL.
- On-premises LLM platforms: model serving with vLLM, an access-controlled model gateway, and retrieval for document chat.
- Model evaluation: reproducible benchmarks for quantized models covering translation quality across 46 language pairs, latency and throughput.
- Efficient NLP research: distilled encoders for 16 languages (TiME), presented at the RethinkingAI Workshop at EurIPS 2025.
Experience
Apr 2026 – Present
Ergon Informatik AG
Machine Learning Engineer · Zurich, Switzerland
- Built reproducible benchmarks for quantized language models, evaluating translation quality across 46 language pairs as well as inference latency, throughput, and robustness.
- Deployed on-premises large language model (LLM) services and an access-controlled model gateway; integrated a document chatbot with multilingual embeddings, reranking, and vector search.
- Extended and deployed a full-stack document translation application with multilingual optical character recognition (OCR), automatic language detection, and streaming results.
PythonKotlinTypeScriptAngularDockerKubernetesvLLMLiteLLMPostgreSQL
2024 – 2025
dreifach.ai
AI Engineer (Working Student) · Remote / Cologne, Germany
- Built LLM-backed tools for insurance workflows: document analysis, internal chat, and retrieval.
- Implemented embedding-based contract comparison, automated ticket tagging, and privacy-aware archival pipelines.
PythonTypeScriptLangChainAzureRAG/AgentsDocker
2024
Karlsruhe Institute of Technology
Data Science Tutor · Karlsruhe, Germany
- Supervised Data Science Lab projects: grading, code reviews, and project guidance.
2021 – 2023
Vector Informatik GmbH
Software Engineer (Working Student) · Karlsruhe, Germany
- Developed Java tooling for the PREEvision CASE suite.
- Designed and shipped a plugin for the propagation rule framework and presented it to the department.
JavaMavenSVNJUnitMockito
Publications
- TiME: Tiny Monolingual Encoders for Efficient NLP Pipelines RethinkingAI Workshop at EurIPS 2025, Copenhagen · arXiv:2512.14645
- A Human-AI Collaboration System for ADHD Assessment from Primary School Reports CHI 2026 Extended Abstracts (CHI EA ’26)
Education
2023 – Apr 2026
Karlsruhe Institute of Technology (KIT)
M.Sc. Computer Science · Karlsruhe, Germany
- Final grade 1.2; master’s thesis graded 1.0.
- Focus: AI, Data Science, and IT Security.
- Research: explainable AI for micro-expression recognition.
2024 – 2025
École Polytechnique Fédérale de Lausanne (EPFL)
Swiss-European Mobility Program, Computer Science · Lausanne, Switzerland
- Machine Learning, Advanced Probability, and Applied Data Analytics.
- Master’s thesis: Tiny Language Models for NLP Pipelines, with Prof. Robert West.
2019 – 2023
Karlsruhe Institute of Technology (KIT)
B.Sc. Information Systems · Karlsruhe, Germany
- Specialization in Software Engineering, Data Science, and Finance.
- Thesis: Hidden Outliers in Manifolds (grade 1.0).
- Exchange semester at Budapest University of Technology and Economics (BME), 2022 – 2023.
Research & Projects
2025
EPFL Data Science Lab
TiME: Tiny Monolingual Encoders for Efficient NLP Pipelines
- Distilled MiniLMv2-based encoders for 16 languages in three sizes each, for deployable NLP pipelines.
- Evaluated on POS tagging, lemmatization, dependency parsing, NER, and QA; measured latency, throughput, and energy per sample.
- Released models and code; presented at the RethinkingAI Workshop at EurIPS 2025 in Copenhagen.
PyTorchTransformersKnowledge Distillation
2025 – Present
Karlsruhe Institute of Technology
Explainable NLP for ADHD Anamnesis
- Fine-tuned encoders and applied supervised fine-tuning and reinforcement learning to Qwen3-4B for criterion-level analysis of German primary school reports.
- Encoder experiments in the paper use focal loss with inverse-frequency class weighting and are compared against open-weight LLMs on macro F1 and class-specific recall.
- Python backend (models, REST API) and React/Node.js frontend with evidence highlighting and clinical metrics (AUC/ROC, sensitivity, specificity).
PythonReactNode.jsXAI
2024
EPFL, with the Swiss AI Center and ETH Zurich
Handwriting Synthesis Pipeline for Dataset Generation
- Generated realistic handwritten images from LaTeX math exercises to stress-test OCR systems and produce multilingual training data.
2023 – 2024
Karlsruhe Institute of Technology
Explainable AI Lab: micro-expression recognition
- Implemented and compared explainable approaches for facial micro-expression recognition.
2023
Karlsruhe Institute of Technology
Hidden Outliers in Manifolds (bachelor thesis)
- Developed methods to generate and detect hidden outliers in high-dimensional data using autoencoders.
PythonAutoencodersAnomaly Detection
Skills
- Languages
- Python, Kotlin, TypeScript, Java, SQL
- ML and LLMs
- PyTorch, Hugging Face Transformers, TensorFlow, vLLM, LiteLLM, LangChain, RAG and agents
- Infrastructure
- Docker, Kubernetes, Linux, Azure, PostgreSQL
- Data
- pandas, NumPy, scikit-learn, Polars
Languages
- German
- Native
- English
- C1
- French
- B2
- Russian
- A2
Service
2020 – 2023
Hadiko Student Dormitory
Resident Speaker · Karlsruhe, Germany
- Represented 100+ residents, moderated weekly meetings, and coordinated community initiatives.