CV

Education

Integrated Master’s in AstrophysicsSeptember 2021 - June 2026

First Class Degree - University of St. Andrews

Highlighted Courses: Machine Learning for Data Analysis; Advanced Data Analysis; Computational Physics/Astrophysics; Cosmology; Observational Astrophysics
Skills: Python modelling (ML: PyTorch, scikit-learn, Hugging Face; Bayesian inference: emcee, nautilus); Fortran Monte-Carlo simulations; experience with Linux
Achievements: Dean’s List in 1st, 2nd, 5th years of university; school prize for best astrophysics research project; school prize for best project poster; invited to Scottish University Physics Alliance annual conference to present research project

Persona Selection Covert Red TeamingSeptember 2026 - Present

SPAR Fall 2026 Research Fellowship Project

Inducing emergent misalignment via covert prompts that shift models off the assistant persona, with white-box probes to measure the shift; mentored by Ben Maltbie (Pivotal Research / MIT).

Timescape Cosmology vs. Dark EnergyJanuary 2026 - May 2026

Masters Project as a part of curriculumAvailable on GitHub

Investigation into performance of inhomogeneous and dynamical dark energy cosmological models.
Constructed data pipelines to combine overlapping catalogues of Supernovae and large-scale galaxy distribution.
Optimised performance of model fitting by 10x using interpolation methods and parallel-processing, enabling improved parameter fitting and modelling with larger datasets.
Tested Markov-Chain Monte Carlo and nested-sampling fitting methods, validating performance against known benchmark models.
Found timescape fits individual datasets well, but struggles to generalise to joint-probe fits, with a discrepancy in inferred void fraction between early- and late-universe fits.
Awarded school prizes for best research project and project poster in astrophysics, preparing paper for publishing with supervisors.

Margnap - AI Text ClassifierJuly 2026 - Ongoing

Independent project to develop ML skillsAvailable on GitHub

Fine-tuned DistilBERT classifier using publicly available multi-domain AI/human dataset from Hugging Face.
Identified in-domain 99.9% accuracy collapsing to random chance (50.05%) out-of-distribution as confounding signals in training data.
Evaluated acceptable FPR/FNR thresholds to balance accuracy with high impact of falsely flagging human-written content as AI, prioritising minimal FPR ≤ 1% at expense of FNR.
Tested performance using OOD data and adversarial perturbations (paraphrasing, word choices, character usage).
Differentiated discrimination failures from calibration errors across domains, recognising where threshold tuning vs. retraining could recover performance.

Elected School President of Physics & AstronomySeptember 2024

Concentrated on holistic school engagement program with more student-staff social initiatives such as continued ’Pints with Professors’ events. Led mid-semester intervention on disrupted module teaching, working alongside student and staff representatives to organise a targeted approach to course-correcting.