Artificial intelligence · Statistics · Law
Research student in statistics & AI at the London School of Economics, working on evaluation methodology, the geometry of space, and statistical methods that take account of the hardware on which they are computed.
Every point of the ordinary sphere lifts to a circle in the three-sphere, and the circles over one line of latitude fill a torus: three latitudes, three linked tori, projected stereographically. Slide to move the latitudes or to change the size, drag to turn.
I am a research student in statistics & AI at the London School of Economics, reading for the MSc Statistics by Research (Finance). My work sits where evaluation methodology, geometry and computation meet. I came to statistics from law, and I still work along the boundary between the two.
Technical research
Other research
Education
Distinction to date. Quantum computation (geometric linear algebra), coursework (PDF); graph learning, statistical inference, financial statistics, financial mathematics in practice (certificate, PDF). Dissertation: Construct Validity in the Evaluation of Causal Reasoning in Language Models. Course representative.
Distinction. Neural networks and deep learning, machine learning, data mining, information retrieval and NLP, risk, statistics, ethics. Course representative.
Financial regulation, securities and banking law, mergers & acquisitions, and the law of international finance.
Qualifying law degree.
Certificates
Good Merit. Designing intelligent agents, advanced algorithms and data structures, systems and networks; research project in reinforcement learning.
Other experience
Talks, posters and workshops
Whether an evaluation measures the construct it claims to. I treat evaluation as a measurement problem: the validity of metrics and benchmarks, the conditions under which their conclusions hold, and how the instruments used to judge models, including models used as judges, should themselves be validated before their verdicts are relied upon.
The choice of geometric space as a modelling decision. Generative models are built in Euclidean space by default; I study the alternatives, with curvature, symmetry and topology treated as parameters to be matched to the structure of the data, and the consequences of that choice for the stochastic processes, metrics and sampling schemes on which a model depends.
Statistics that accounts for the machine it runs on. The performance of an estimator is conventionally judged by statistical error alone; I study the trade offs that arise once computational cost, memory and parallel hardware are treated as part of the problem, and the conditions under which the preferred method changes with the environment. I lead the LSE HPC team, which won one of the two main challenges at the CIUK Cluster Challenge and placed third overall, and I have spoken at Imperial’s HPC-SIG on building a successful HPC educational initiative.
Generative modelling and evaluation for the life sciences. Diffusion processes on non-Euclidean spaces for molecular conformation; the construction and optimisation of genome scale metabolic models; the evaluation of clinical prediction models; and metrics for graph structured, multimodal health data held within a national secure research environment.
Law as text, and text as evidence. I undertake a broad range of research in legal settings; my particular expertise lies in evaluation and benchmarking, causality in reasoning, linguistic time series, and sentiment and stance analysis, applied to legal, regulatory and political corpora. Beneath all of it is the modelling of geometric space, which is what allows causal and temporal analysis to be constructed at a theoretical level; the papers themselves rest on rigorous experimentation and are firmly grounded in the empirical method.
Model risk, systemic risk and financial data. The reconstruction of financial networks and the sensitivity of supervisory conclusions to the method of reconstruction; the estimation of risk under computational constraints; and the extraction of structured information from financial data with language models, at scale.
Reviewing
Reviewer for FinNLP@EMNLP, ICAIF and JUDGe@NeurIPS.
Roles
Committees
Publications and talks on teaching
Standards
Member of the working group for IEEE P3722, a standard for ethical generative AI in education, from 2025 to 2026.
Team lead of No Human Intelligence, one of 36 finalists out of 336. We built the Herding Scenario Lab, a decision first sandbox for financial stability supervisors: it stress tests AI market surveillance tools and intervention rules against synthetic markets of herding AI trading agents, calibrated from the regulator’s own data, and certifies each tool on the supervisory decision it exists to inform rather than on aggregate fidelity, with a conformal false alert guarantee, a signed replayable ledger, two human gates and a red team. Its central finding, that ranking tools by fidelity nearly reverses ranking them by decision accuracy, is the subject of Reconstructing Systemic Risk: How Synthetic Network Generators Shape Banking Stress Tests, accepted at ICAIF 2026. The hackathon was featured in the Financial Times in Gillian Tett’s article, ‘How can we tackle AI’s financial risks?’, on 25 September 2026. Finalist certificate (PDF).
Clinical track: can an AI be trusted in intensive care? On a synthetic cohort of 5,000 post surgical ICU patients spanning tabular, clinical note and hourly vital sign modalities, we built a multimodal fusion model and a tabular baseline for deterioration within 24 hours, and designed the evaluation around trust: missingness that is not at random and concentrates in the sickest patients, calibration, subgroup equity across age and surgery type, and profiling of who is missed and who triggers false alarms, reported as a model evidence card and safety report that a hospital safety committee could act on. Best Multimodal Idea, in a field of PhD students and postdocs. Code on GitHub · Award certificate (PDF).
Pacifico Biolabs main track. A genome scale metabolic model pipeline for physiological strain optimisation, sponsored by the CHAI Hub. Second place certificate (PDF).
High performance computing. Team lead of the LSE team, fully funded by CIUK; winners of one of the two main challenges and third overall in the finals.
Team lead. Designed and built a tool predicting the value and risk of legal settlements from case data, for use in litigation strategy; sponsored by Clifford Chance.
Daria Godorozha