MPhil Advanced Computer Science @ Cambridge · First-Class BSc @ St Andrews
Research scientist / research engineer focused on mechanistic interpretability, LLM evaluation, and ML systems.
🗄️ My CV as a 3D physics junk drawer → — ten things I've built, dumped in a drawer. Grab one and read it.
Current research: mechanistic interpretability, readout-side analysis, and RL for reasoning models.
Previously: ML infrastructure at Amazon Alexa-AI · Browser-based LLM inference with WebGPU (3K MAU, zero server-side inference cost).
Interests: AI safety · scalable oversight · empirical ML research · reproducible ML systems
Stack: Python · PyTorch · TransformerLens · JAX · AWS · Docker
- Sparse Readout Prism — arXiv:2609.01936. Explains logit-lens scores as sums of sparse readout-feature contributions, with code and pretrained dictionaries.
- vigil — real-time TUI for monitoring rented cloud GPU training instances.
- Browser-based LLM inference — WebGPU/WebLLM system serving local model inference in the browser with no backend inference cost.
- Dolphin acoustics classification — ML experiments for species classification from whistle-contour features.
- Mechanistic interpretability of language models
- Readout/unembedding dynamics and sparse representations
- RL, evaluation, and scalable oversight for reasoning models
- Infrastructure for reproducible ML experimentation