About
Machine learning engineer and scientist, based in London. Knowledge graphs, and the evaluation problems they create.
I work on turning fragmented, unstructured sources into graphs that can be queried and trusted, and on the machine learning that ranks and predicts over them. In practice that means ontology and schema design, entity resolution across sources that disagree with each other, relationship extraction from documents, provenance modelling, and graph learning for link prediction.
Now
I am sole engineer on a literature-to-knowledge-graph platform for the Koslicki Lab at Pennsylvania State University, contracted from London. The pipeline screens roughly 2.9 million candidate records down to a working corpus of 100,000 documents, extracts structured entities and relationships from the ~90,000 available in full text, and delivers a Neo4j graph in which every edge carries the source document and the exact text it derived from. A LightGCN-style graph neural network ranks interventions by the quality of the evidence behind them, with post-hoc attribution over its output.
I also hold a postdoctoral position at Evotec, working on record linkage and data quality for a commercial predictive service, and I am a Technical Champion in the NIH-funded NCATS Biomedical Data Translator consortium, which operates one of the largest federated knowledge graphs in production.
Before
My PhD, at the University of Sheffield and co-funded by Evotec, was on interpretable completion of biomedical knowledge graphs. Rather than embeddings, it uses counts of typed paths through the graph, so every prediction traces back to named entities and relations. Its main contribution is Inverse Node Frequency, a feature transformation adapted from inverse document frequency that corrects for degree bias: the tendency of path-based methods to predict well on densely connected entities and badly on sparse ones, which are usually the ones worth predicting. It outperforms the field-standard Degree-Weighted Path Count on recall for low-connectivity entities, and holds that advantage on an external test set never previously used for the task.
A result I am fonder of than the headline one: a set of features that appeared highly predictive turned out, under analysis, to be a topological artefact, and the affected results were discounted. The interpretable method made that visible. An embedding model would have shipped the number.
Before research I spent four years in investment banking and corporate finance in London, Munich and Cologne, covering energy, utilities and renewables, and mergers and acquisitions. I have an MSc in Data Science and Artificial Intelligence from Queen Mary University of London and a BA in Economics from Dartmouth College. I speak German.