📍 Trier, Germany  |  🎓 M.Sc. AI @ Hochschule Trier (June 2026)  |  🏢 Open to EU Relocation (No Sponsorship Required)

I build evaluation infrastructure and production-grade architectures for language models and multi-agent systems. My core thesis is empirical: standard metrics and model confidence scores frequently obscure actual internal behavior.

By bridging mechanistic interpretability with scalable MLOps and Azure cloud infrastructure, I engineer reproducible evaluation pipelines, high-performance RAG architectures, and agentic workflows (LangGraph, Model Context Protocol) designed for production reliability and regulatory compliance.

• Languages: Python, SQL
• ML & LLM Ecosystem: PyTorch, TransformerLens, Hugging Face, scikit-learn, XGBoost, LightGBM, SHAP, Optuna
• Agentic Frameworks & RAG: LangGraph, Model Context Protocol (MCP), Qdrant, LlamaIndex, FAISS, Hybrid Search (Vector + BM25)
• Cloud & MLOps: Azure Machine Learning (SDK v2), Azure OpenAI Service, Azure AI Search, MLflow, FastAPI, Docker, Docker Compose, GitHub Actions, CI/CD

Python PyTorch TransformerLens PyPI Docker FastAPI

• Created and maintain glassbox-mech-interp (v4.5.0 on PyPI), an open-source toolkit implementing 21 mathematical frameworks for causal circuit discovery via activation patching.
• Benchmarked 15 to 37 times faster than ACDC on GPT-2 Small (1.2s on CPU).
• Extended architecture adapters to support Grouped-Query Attention (GQA) and RMSNorm models (Llama-3, Mistral, Phi-3, Gemma).
• Automated compliance reporting, emitting structured JSON for EU AI Act Annex IV documentation.

LangGraph (0.3) Model Context Protocol (MCP) Qdrant LlamaIndex FastAPI Celery

• Orchestrated a multi-agent system utilizing LangGraph to coordinate 3 Model Context Protocol (MCP) servers (Postgres, document, notification).
• Engineered hybrid RAG pipelines over Qdrant using LlamaIndex with a FastAPI backend streaming over WebSockets to a Next.js UI.
• Integrated enterprise observability using Langfuse, OpenTelemetry, Prometheus, Grafana, and Sentry under a robust pytest-asyncio CI/CD pipeline.

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

• Developed a document Q&A platform combining vector embeddings and keyword matching (BM25) for high-accuracy hybrid retrieval.
• Implemented custom chunking and ingestion pipelines over Azure Blob Storage using text-embedding-3-small.

Azure ML SDK v2 MLflow scikit-learn Docker

• Designed an automated 4-stage pipeline (data preparation, training, evaluation, model registration) running on auto-scaling compute clusters that scale to zero idle nodes.
• Integrated MLflow tracking and the Azure ML Model Registry for complete experiment reproducibility and version control.

• Causal Attribution for Agentic Decisions: Estimators, Coupling, and a Traceability Specification (Sep 2026)https://doi.org/10.48550/arXiv.2609.06445 Preprint evaluating total and direct effect estimators for agent trajectories and establishing a 12-point traceability specification for high-risk AI.
• Explanation Multiplicity: Circuit-Level Interpretability Evidence Does Not Survive Defensible Analytic Variation (Aug 2026) – arXiv:2608.13754.

Proved that compliance-related circuit claims flip across 73.2% of specification pairs in a pre-registered multiverse analysis.

• Explainable AI for LLMs: A Causally Grounded Pipeline (Feb 2026) – arXiv:2603.09988. Established a near-zero correlation ($r = 0.009$) between model confidence and explanation faithfulness, proving confidence cannot serve as an evaluation proxy.

• 🎓 Education: M.Sc. in Interdisciplinary Engineering (AI/ML) from Hochschule Trier (Thesis graded 1.0, Overall 1.8).
• 🎯 Status: Available full-time immediately for Machine Learning Engineer, AI Engineer, or LLM Evaluation Engineer positions across the EU (German residence permit, no sponsorship required).

Building robust systems that open up models and make evaluation verifiable.