Prior to completing a PhD in Reinforcement Learning (RL) in 2024, I worked for 18+ years as a highly respected enterprise developer-analyst in multi-national companies. I therefore possess a deep understanding and considerable experience with enterprise project environments and key business processes. I am a strong communicator, having represented development as technical team lead within cross-functional projects. I am results-oriented, have worked successfully within the agile framework and coming from a pharmaceutical background I care about the quality and societal impact of my work.
My PhD equipped me with a strong background in Machine Learning (ML) methods with experience implementing and extending ML, especially RL and DRL, algorithms. I am comfortable analysing data, debugging machine learning algorithms, building tools for analysis and logging while making use of the latest features in existing frameworks and tools. Having previously worked extensively with high profile production systems, I code in a production-aware manner. I am a strong advocate of good design and development practices that I encourage in all teams I work with.
During my PhD I was fortunate to work with a diverse range of people, some of whom I now collaborate with on on-going research initiatives. I have both tutored and mentored students, run workshops on coding and deep dives and given talks on ML and RL.
Current work, projects and initiatives
I have just completed the Blue Dot AI Safety Fundamentals Alignment course (Oct 2024-Feb 2025) [certificate]. It is an excellent well structured course providing a structured view of the AI Safety landscape. I started a blog to document my work in AI safety. My course project looked at training Constitutional AI (CAI) for alternative constitutions that reflected a country or culture more strongly. Specifically I considered how an LLM model would behave if the alignment principles were extracted from the South African (SA) constitution.
The SA constitution includes a strong focus on historical injustices and inequalities, reconciliation and recovery measures and a very explicit effort to prioritise fairness while avoiding biases. A comparison with the constitution used to train Claude shows several differences and raises questions about how to generate the fine-tuning dataset, what impact the new constitution would have and how evaluations should be adapted for very different social and cultural settings.
From January to April 2025 I am working on an AI Safety Scientist project run by the AI Safety Camp. Sakana's AI Scientist currently generates hypotheses, modifies code templates, runs experiments and generates papers for peer review. We are adding scaffolding tools relevant to AI Safety including synthetic dataset generation and evaluation, using the AISI's Inspect framework to generate tasks, initially using pre-existing solvers and scorers but in future possibly adding this as a template for expansion/modification by the AIS scientist. Additionally we are adding AI control protocol scaffolding for automated protocol creation and testing.
APART Hackathon (Mar 2025): Feature based analysis of cooperation-relevant behaviour in prisoner's dilemma .We analysed multi-Agent feature steering and analysis in a multi-agent setting. Our team is continuing to extend the preliminary experiments from Prisoner’s Dilemma to more real-world scenarios using Concordia.
APART Hackathon (Mar 2025): Participating in the AI Control hackathon from APART research, developing and testing new protocols for the DTT (defer-to-trusted) setting. Code is implemented using the Control Arena framework from AISI and will be released after the hackathon.
I am a member of the steering committee of the TVAI (Thames Valley AI). This is an initiative driven by academia, industry and the local community to generate awareness and innovation in AI. We are running events in 2025 so if you would like to speak at one of our events please do reach out (LinkedIn).