Master’s student in Computer Science, Saarland University.
Junior AI safety researcher.
juliuskamp2000 [at] gmail [dot] com · GitHub · LinkedIn · Saarbrücken, Germany
Education
Master’s thesis (in progress): Automated Debugger for AI agents.
Exploratory Data Analysis group (Prof. Jilles Vreeken) at CISPA, advisor: Nils Walter.
Bachelor’s thesis: Discovering Fully Oriented Causal Networks from Discrete Observational Data
Exploratory Data Analysis group (Prof. Jilles Vreeken) at CISPA, advisor: Osman Mian.
Experience
Exploratory Data Analysis group.
Research on mechanistic interpretability in large language models.
Modeling and Simulation group.
Built and maintained ProcGrid Traffic Gym (PGTG), a Gymnasium-compatible reinforcement learning environmen. PyPI · Docs · GitHub
AI Safety
Two-week full-time intensive at Trajan House, Oxford. Compressed ARENA syllabus.
Capstone: replicated and extended the Intentional Control of Internal States experiment from Anthropic’s introspection paper on Gemma 3 27B. Write-up · Code
Ten-week program on evaluations and interpretability. Ran the application process, speaker outreach, and some of the weekly sessions.
Scoped and ran a project on Natural Language Autoencoder–based measurement of evaluation awareness.
Project on finding subspace circuits in transformers, applying the Subspace Partition method to toy transformer models.
Skills
- ML / interpretability: PyTorch, TransformerLens, mechanistic interpretability (probes, activation steering, SAEs, circuit analysis), evaluations, reinforcement learning
- Engineering: Python, Git, HuggingFace transformers, remote GPU workflows
- Also familiar with: C, C++, HTML, JavaScript
- Languages: German (native), English C1 (UNIcert III)