How confident should this prediction be?
Bayesian kernel methods and deep Gaussian processes that give language models calibrated confidence, so low-confidence cases go to a person instead of through as settled.
Open to research and applied scientist roles
Machine learning that knows when it doesn't know.
PhD researcher in trustworthy and responsible AI at the University of Southampton. Primary researcher on an Alan Turing Institute grant.
About
I build machine learning systems that can tell when their output should be trusted. My work combines Bayesian uncertainty for language models, explanations that are tested for faithfulness rather than assumed, and privacy-preserving adaptation for settings where data cannot be pooled. I have applied it to clinical and legal text, online harms and under-represented languages.
I am the primary researcher on PICASO, a competitively awarded Alan Turing Institute grant, and I led the TrustWeave team in the United Nations Development Programme's AI Trust and Safety Re-Imagination Programme. I am open to research scientist and applied scientist roles in academia and industry.
Research questions
Bayesian kernel methods and deep Gaussian processes that give language models calibrated confidence, so low-confidence cases go to a person instead of through as settled.
Span-level rationales for harmful content, evaluated for plausibility and faithfulness, and stress-tested under adversarial variation and shift over time.
Lightweight federated prompt tuning that adapts language models across clients and languages without centralising sensitive data.
Uncertainty-aware knowledge graph embeddings on conceptual spaces, where the geometry carries both meaning and confidence, plus an open Wikidata-derived benchmark.
Featured · CIKM 2026, Rome
Knowledge graph embeddings that represent each entity as a region in a conceptual space, so a predicted fact comes with a calibrated measure of how far it can be relied on. Released with Defence-Wikidata, an open benchmark built from a full Wikidata dump.
News
Presenting PROCS: Uncertainty-Aware Knowledge Graph Embeddings via Conceptual Spaces at CIKM 2026 in Rome.
Presenting at the UK AI Research Symposium (UKAIRS) 2026 in Edinburgh.
PROCS, with Steven Schockaert, accepted at CIKM 2026.
Invited talk at the Learning and Reasoning Seminar, Cardiff University.
Won two University of Southampton Doctoral College Director's Awards, for Public Engagement and Outreach, and Knowledge Exchange and Enterprise.
Presented Towards Explainable Hate Speech Detection in Roman Urdu at IJCNN 2026 (IEEE WCCI).
Started as primary researcher on PICASO, an Alan Turing Institute grant on probabilistic conceptual spaces for sensemaking under uncertainty.
Led research team TrustWeave in the UN AI Trust and Safety Re-Imagination Programme.
Publications
PROCS: Uncertainty-Aware Knowledge Graph Embeddings via Conceptual Spaces
Contact
Research collaborations, talks, or research and applied scientist roles.