I am a Senior Data Analyst and Independent AI Researcher based in Berlin, Germany.
With over 10 years of experience (most recently at Klarna), I specialize in the intersection of rigorous statistical methodology, data engineering, and machine learning. My current focus is on AI Safety, specifically investigating alignment failures in Large Language Models (LLMs) and building robust evaluation pipelines.
I am currently seeking opportunities that blend my background in data analytics at scale with my passion for AI research and engineering.
The Devil in the Details: Emergent Misalignment, Format and Coherence in Open-Weights LLMs
Craig Dickson (2025)
I conducted an independent study replicating and extending findings on "emergent misalignment" across 9 modern open-weights models (Gemma 3, Qwen 3). The research identified a critical vulnerability where JSON output constraints double misalignment rates compared to natural language.
| Resource | Description |
|---|---|
| 📄 Read the Paper (arXiv) | Full methodology, statistical analysis, and findings. |
| 💻 View the Code | End-to-end pipeline: LoRA fine-tuning (unsloth), vLLM inference, and LLM-as-judge evaluation. |
| 📊 HuggingFace Dataset | The full dataset of over 50,000 model responses generated during the study. |
| ✍️ Blog: The JSON Trap | An accessible breakdown of why structured outputs make models less safe. |
Data Engineering & Analytics
AI & Machine Learning
I'm continuing to expand my work on emergent misalignment, with plans to use Sparse Autoencoders (like Google's Gemma Scope 2) to empirically test hypotheses from my paper The Devil in the Details. I'm very interested in collaborating on research with anyone who shares a commitment to building AI safely and deploying it in ways that help people.
I'm also exploring how modern AI tools can enhance analytics performance and productivity in professional workflows. This is something I am working on intensively in my professional work right now.
The Guardrail - AI Safety Paper Aggregator | Live Site
JavaScript, LLM Integration
A daily-updated aggregator of AI safety research from arXiv. Features AI-powered filtering and categorization across 10 safety domains including alignment, interpretability, RLHF, robustness, and agent safety. Currently tracking 1,300+ papers with relevance scoring and editor's picks.
Lumina - Article to Audio Converter | Live Demo
React, TypeScript, Google Gemini TTS API
A web application that transforms articles and text into natural-sounding audio using Google Gemini's text-to-speech capabilities. Features voice selection, playback speed control, and paragraph-level highlighting with click-to-jump navigation.
Personal Research Librarian
Python, LanceDB, sentence-transformers, Google Gemini
A CLI tool for building a private RAG (Retrieval-Augmented Generation) database from your personal documents. Ingest PDFs, notes, and documents to create a local vector index, then query your knowledge base using natural language with grounded answers and citations.
Alexa-to-Gemini Bridge
Python, Flask, Google Gemini API
A middleware application that upgrades legacy Amazon Alexa hardware by routing queries through Google's Gemini 2.0 Flash Thinking model. This enables complex reasoning and natural conversation on older smart speakers.
Universal AI Chatbot Interface
Streamlit, Multi-Provider API Integration
A versatile chat platform supporting hot-swappable providers including Google Gemini 2.0 Flash Thinking, Deepseek (r1 & Chat), OpenAI GPT-4o, and Anthropic Claude.
Don't hesitate to drop me a line if you have any questions about my research, or want to discuss working together.