Researcher · b. 2006
Nineteen-year-old researcher working across machine learning, medicine and materials science. Currently investigating data-efficient interatomic potentials at the University of Cambridge, and co-developing a device for the early detection of postpartum haemorrhage with Patch.
scroll
Data-efficient interatomic potentials
Investigating how to train machine-learned interatomic potentials with far less data, using datasets of vacancies, interstitials and substitutional defects, benchmarked against Density Functional Theory. The goal: predict defect formation energies accurately while cutting the computational cost of atomistic materials research.
Early detection of postpartum haemorrhage
One of 36 selected from 500+ applicants. Co-developing BLOOM, a low-cost, offline device that uses transmission spectroscopy to track haemoglobin loss during caesarean delivery, providing an earlier warning than visual estimation and designed for low-resource and humanitarian settings.
Multimodal deep learning with clinician-facing explanations.
A multimodal explainable-AI system that uses MRI and clinical data to predict clinically relevant molecular and outcome variables in adult diffuse glioma. The project also includes treatment-effect analysis, a clinician-facing interface, and an exploratory health-economic assessment.
Won Stripe Young Scientist & Technologist of the Year 2026. Presented at Stryker, Cubic and Stripe Sessions; selected for EUCY 2026.
Transmission spectroscopy for real-time haemoglobin tracking.
BLOOM analyses fluid collected during caesarean delivery using transmission spectroscopy, tracking the rate of haemoglobin loss to provide an earlier warning than visual estimation. The device is being designed for low-resource and humanitarian settings: offline, low-cost, and robust.
In active development with the PATCH 2026 cohort. 36 participants chosen from more than 500 applicants.
Benchmarking OCR and vision-language models against noisy clinical data.
Researched and benchmarked OCR and zero-shot vision-language approaches for extracting REF, LOT and NDC identifiers from noisy clinical inventory images and records. When the available labelled data proved insufficient to train a reliable custom model, I recommended an API-based pipeline instead.
Metadata-extraction accuracy increased from approximately 30% to over 90% in internal testing. Subsequently offered a paid summer role.