About
Hi! I'm an AI/ML Engineer with a background in statistics. I like to start with basic questions: what is being measured, what is being assumed, and where does the model become overconfident?
During my master’s at Saarland University, I studied temperature-scaled surprisal and eye-tracking data. The question was simple on the surface: can softened model probabilities explain human reading times better? The harder part was tokenization. Once a word is split into subwords, the signal changes depending on where the subword sits. A preprocessing step suddenly becomes part of the interpretation.
I have worked across research and engineering: arithmetic reasoning in language models, robust learning under distribution shifts, neural policy verification, recommender systems, forecasting, and AI workflows inside production software.
I like the space between a clean research question and a system people can actually use. That space is usually where the interesting problems appear.
Experience
Machine Learning Engineer · AICU GmbH
Built production ML features for a high-paced SaaS startup, covering recommendation workflows, forecasting pipelines, AI-agent interactions, and client-specific automation projects. Worked across the full path from problem definition to deployment: translating client requirements, designing model and data pipelines, integrating them into backend services, and monitoring them in production.
- Recommendation Systems
- RAG
- AI Agents
- Qdrant
- AWS
- Django
- Docker
Research Assistant · Saarland University · CompLing
Can LLMs subtract numbers?Built evaluation pipelines for controlled arithmetic experiments in language models. The work focused on cases where model outputs fail even when useful information appears to be represented internally, especially around subtraction and sign prediction. Ran experiments across model families, digit lengths, and controlled train-test splits to separate memorization from generalization.
- LLMs
- Reasoning
- Evaluation
Research Assistant · CISPA
Confounder-Agnostic Learning: Enhancing Robustness with Projected Gradient DescentWorked on robust machine learning under distribution shifts. The project used a modified min-max PGD setup to reduce confounding effects without relying on known anchor variables. The work focused on how models behave when spurious correlations change between training and test distributions.
- Robust ML
- PGD
- Distribution Shift
Research Assistant · Foundations of Artificial Intelligence Group
Ran empirical evaluations for safety verification of neural action policies in symbolic planning environments. Designed environment variants, trained reinforcement-learning policies, and tested how environment choices affected verification complexity and policy safety. The role connected learning-based policies with formal methods, where high task performance alone was not enough.
- Reinforcement Learning
- Planning
- Verification
- CEGAR
- Python
Projects
VisionRAG from Scratch
Built a visual PDF question-answering system that indexes pages as images instead of flattening them into plain text. The pipeline uses visual page embeddings, Qdrant, a retrieve-then-answer flow, and bounding-box style evidence marking so the answer can point back to where it came from on the page.
- ColPali
- VisionRAG
- Document AI
- Retrieval
- VLMs
Temperature-scaled Subword Surprisal
Studied how language-model surprisal relates to human reading times when words are split into subword tokens. The work tested temperature scaling with GPT-2 and mGPT on eye-tracking data, then compared what changes at word level, subword level, and across languages.
- Surprisal
- Eye-tracking
- Tokenization
- mGPT
- Mixed-effects models
CLIP Text-to-Image Search
Built a small image search system using a contrastive dual-encoder setup. Image and text representations were trained into a shared space, then indexed for fast lookup so a text query could retrieve matching images rather than only classify them.
- CLIP
- FAISS
- ResNet
- Transformer
- InfoNCE
Writing
VisionRAG from Scratch
A technical article about implementing VisionRAG using visual document retrieval instead of only OCR-based text extraction.
Read article →Contact
I am open to ML engineering and research-facing roles.
Email is the easiest way to reach me.
Get in touch