Alexander F. Spies

AI Safety Researcher (PhD, Imperial College London) specializing in mechanistic interpretability and causal world models, with hands-on experience shipping robust LLM systems and building eval/finetuning/tool-use pipelines. Published on maze-task world models & transformer internals; led research teams and developed safety-relevant tooling. Seeking to work on evaluations and interpretability techniques for frontier models.

Thesis: Interpretable Representations in Artificial Neural Networks

• Improving representations in Object-Centric Learning and reasoning
• Mechanistic analysis of vision & language transformers
• Advisors – Prof. Alessandra Russo & Prof. Michael Shanahan

• Thesis: Learning World Models in the Animal-AI Environment
• Independent project: Neurosymbolic Learning & Neurally Weighted dSILP

• Completed graduate-level courses as an undergrad, alongside research

• Thesis: AI for the Automated Diagnosis of Atrial Fibrillation

Research engineer working on production LLM systems and agentic pipelines (promoted from intern to full-time in June 2025).

• Shipped production LLM services (10k+ weekly queries) for low-resource language support; hardened prompts, data paths, and fallback behaviors for robust performance across diverse use cases.
• Built finetuning/eval infra: UnSloth & SageMaker training, W&B sweeps, vLLM serving; standardized experiment tracking → faster ablations + reproducible benchmarks.
• Developing embedded agentic pipelines for multi-turn code generation, repo understanding, and retrieval + tool-use.

Led independent research groups on mechanistic interpretability as "model organisms of misalignment."

• Defined agenda on mechanistic interpretability for maze-solving LMs; trained transformers and Sparse Autoencoders, managed 9 researchers across 2 projects → 2 workshop papers + best-poster award.
• Built SAE-based analysis pipelines to study internal circuits & world-model structure; results demonstrated causal world models emerging in maze-solving transformers.
• Research artifacts, figures, and code published in workshops and available on GitHub/personal site.

Mechanistic analysis of Transformers trained on maze‑solving tasks.

Investigated non‑local thresholds in pixel detectors; co‑authored JINST publication.

Exclusion analysis of Higgs decay channels in MSSM.

• Provided technical guidance on AI Safety Research to 8+ Research Fellows

• NeurIPS, ICLR, ICML, AAAI, UAI, Artificial Intelligence (Journal); MATS program applications

• Led technical coursework for Deep Learning, ML Math, Data Structures & Algorithms, and Python
• Engineered GPU-backed autograding pipeline for 120+ students using Otter Grader and Paperspace