📍 Trier, Germany  |  🎓 MSc AI @ Hochschule Trier (May 2026)  |  🏢 ML Engineer @ Basemotion AI, Berlin

I build evaluation infrastructure for language models, not dashboards that look good in demos, but pipelines that surface what metrics actually measure versus what they claim to measure.

My core thesis: confidence scores are lying to you. I proved it with a near-zero correlation (r = 0.009) between model confidence and internal reasoning faithfulness. That finding came from combining activation patching, causal circuit analysis, and a reproducible benchmarking framework I built from scratch.

Currently writing my MSc thesis on explainable AI for LLMs with causally grounded natural language explanations, while working as an ML Engineer at Basemotion AI in Berlin on production LLM evaluation systems.

Python PyTorch TransformerLens arXiv PyPI Hugging Face

The project that came out of a direct question: can we tell, causally, which parts of GPT-2 drove a specific prediction?

• Built a causal circuit discovery engine that answers that question in 1.2s on CPU using 3 forward passes, 37× faster than the ACDC baseline
• Quantified r = 0.009 correlation between model confidence and internal reasoning faithfulness, a result with direct implications for EU AI Act compliance
• Automated generation of all 9 required EU AI Act Annex IV sections from a single function call. Structured JSON output ready for GRC system import
• Published on arXiv (2603.09988), deployed a live Hugging Face demo, and shipped to PyPI with a CLI + 76 automated tests

Compliance teams can audit any model in under a minute with zero infrastructure setup.

Azure OpenAI Azure AI Search FastAPI Streamlit GPT-4o-mini

Document Q&A system with source citations built on Azure's full AI stack.

• Hybrid search combining vector embeddings + keyword matching for semantically-aware retrieval
• Document ingestion pipeline with 512-token chunking and text-embedding-3-small embeddings
• FastAPI backend + Streamlit frontend with streaming responses for real-time answer generation

Azure Machine Learning MLflow scikit-learn Azure ML SDK v2

Automated 4-step ML pipeline: data prep → training → evaluation → model registration.

• 74% test accuracy, 80% F1, 87% AUC-ROC on heart disease prediction (200-record held-out test set)
• Auto-scaling compute with minimum zero nodes. Clusters shut down automatically when idle
• MLflow tracking + Azure ML Model Registry for full experiment reproducibility and version rollback

Explanation Multiplicity: Circuit-Level Interpretability Evidence Does Not Survive Defensible Analytic Variation

The EU AI Act requires providers of high-risk systems to file technical documentation describing how the system reaches its decisions. Mechanistic interpretability is the obvious source of such evidence, and circuit discovery is its most developed instrument. We ask whether that evidence survives the condition under which it would be relied upon: two competent analysts, the same system, the same tool, different defensible settings. We pre-registered a crossed grid of seven analytic axes, every level taken from a published implementation, and mapped each discovered circuit through a deterministic claim map to a structured Annex IV statement. Across 15,840 pre-registered specifications on GPT-2 small and the indirect object identification task, of which 7,561 produced a claim, the derived statement flips across 73.2% of specification pairs (95% CI 0.725 to 0.738) and the modal claim commands 41.1% of the space. The evidence fails a filability criterion at every tolerance a conformity assessment body would plausibly accept. Standardising the single most influential choice, the evaluation metric, leaves the flip rate at 59.4%. Removing circuit size from the claim entirely and holding it fixed leaves 27.1% (95% CI 0.255 to 0.286), still above the pre-registered threshold. The circuits underlying these claims are structurally near-disjoint, median pairwise Jaccard overlap 4%, and functionally uncorrelated at Cohen's kappa 0.015, so the instability is not one mechanism described in different words. We give the filability criterion as a standalone protocol, and we report that one of the seven documented discovery objectives does not execute at all on the library's own canonical task. The study covers one model and one task, and whether the conclusion holds at scale is untested.

Explainable AI for LLMs: Causally Grounded Mechanistic Interpretability for LLMs with Faithful Natural-Language Explanations

The core finding: traditional attention-based metrics miss 39% of prediction behavior. Ground truth established via 100% sufficiency scoring using activation patching and causal circuit analysis. The pipeline converts technical circuit data into structured natural language explanations validated against ERASER metrics.

Languages: Python, SQL

ML / Research: PyTorch, TransformerLens, HuggingFace, scikit-learn, NumPy, Pandas

Cloud / Infra: Azure Machine Learning, MLflow, Docker, REST APIs, FastAPI, GitHub Actions, CI/CD

Core Expertise: Mechanistic Interpretability · Activation Patching · Transformer Architecture · LLM Evaluation Methodology · Causal Analysis · Python Package Development (PyPI) · Prompt Engineering

• 📝 MSc Thesis. Mechanistic interpretability of LLMs with causally grounded explanations
• 🏢 ML Engineer @ One75 Labs. Production LLM evaluation infrastructure, Berlin
• 🎯 Open to. ML Engineer / AI Researcher roles in the EU (post-graduation, May 2026)

I don't just run models. I open them up and see what's actually going on inside.