Edward Stevinson
Hi, I’m Ed, a research scientist and PhD candidate at the CIRCLE group at Imperial College London, where I work on mechanistic interpretability and adversarial robustness. My research centres on understanding the representation geometry of neural networks and how this shapes adversarial vulnerability.
Find me on Twitter, Google Scholar, Github and LinkedIn. Please reach out by email if you want to talk about any research!
work
Research Scientist
Developed an AI safety method to elicit undesirable LLM behaviours through fluent prompt modifications optimised to maximally activate target SAE features. Our paper was accepted at ICLR 2026.
Research Scientist
As code owner of one of the company's neural-network verification toolkits and its certified-training toolkit, I turned research code into production-grade software for certifying and training robust models. Unlocked verification of time-series models, opening new use cases, by implementing support for softmax and 1D convolutional layers.
Research Scientist
Research conducted during an internship focused on leveraging knowledge graphs to update word embeddings in the scientific domain concluding with presenting our paper at an AACL workshop on information extraction.
Senior Software Engineer
Napo Pet Insurance · Contract
Contract python backend developer for the fast growing pet insurance startup. A member of the growth team who helped grow the customer base from 1k to 10k over six months.
Lead Data Scientist
MyUpside · Full-time
Co-founded Upside, a workflow analytics platform that surfaces key metrics and trends to give software teams more autonomy. Built the platform's data and analytics stack end-to-end: a Python analytics engine computing its core metrics, Dagster data-science pipelines that turn the engine's output into the platform's models, and a Cube-powered analytics API driving the customer-facing visualisations.
Software Engineer
Eigen Technologies · Full-time
Eigen is an AI company automating the extraction and classification of information from legal documents. Built the company's custom ML lifecycle management system to reduce the cost of model development, and developed a tool to automate the sectioning of legal documents into a hierarchical representation of their structure.
Software Developer
IBM · Full-time
Server-side Java developer at IBM (i2), building the web-app version of an industry-leading investigative analysis platform used by law enforcement and intelligence agencies.
studies
PhD, Artificial Intelligence
Mechanistic interpretability and neural network robustness verification.
MSc, Artificial Intelligence
Intelligent robotics specialism, with courses in machine learning, probabilistic modelling, reinforcement learning and algorithmic game theory.
MEng, Engineering, Economics and Management
Engineering specialisms in information engineering, machine learning and software; Economics and Management specialisms in strategy and game theory.