Tristan Hodgson

Education

Mmath in Mathematics, University of Oxford

(2024-2028, Expected)

High School

(2018-2024)

Machine Learning Intern, Harper AI

(One Week, Summer 2026)

Developed an XGBoost model identifying refinancing opportunities across 80,000 UK property developers, achieving 88% precision on commercial lead generation
Optimised model performance using RFECV feature selection, Optuna hyperparameter optimisation, and class weighting for heavily imbalanced data
Interviewed customers to refine product requirements, implementing SHAP explanations in response to demand for interpretable model outputs
Engineered point-in-time SQL feature extraction for financial, governance, and group-structure data, preventing temporal data leakage

IT Officer, Trinity College JCR

(Oct 2025-)

Led the migration of the JCR website from a legacy Wix setup to the University-wide Fresco platform, reducing annual IT spend by 94% through platform consolidation
Coordinated approval, requirements, and timelines across multiple College and University teams
Audited and rewrote outdated site content in collaboration with JCR members, restructuring the site as a clear resource for prospective and current students

Process & Data Systems Intern, It's Our Planet Too

(One Week, Summer 2025)

Redesigned operational workflows, replacing legacy spreadsheets using Power Query, User forms, and Pivot Tables
Developed dynamic dashboards tailored to different tasks, simplifying the management of operational data
Created documentation to enable users to understand and maintain the new system

Performance Modelling of In-Database Sparse Matrix Multiplication

Write-up | GitHub

Modelled performance trade-offs between in-database and client-side matrix multiplication under compute, sparsity, and bandwidth constraints
Benchmarked performance to validate asymptotic runtime predictions across matrices of varying size and sparsity
Derived and validated the sparsity threshold function where in-database outperforms client-side computation to inform system architecture decisions

Known Prefix Neural Cryptanalysis with seq2seq Models

Write-up | GitHub

Implemented LSTM seq2seq models in PyTorch to perform neural cryptanalysis of Caesar and substitution ciphers on natural-language sequences
Extended prior neural cryptanalysis work by introducing a known prefix, reducing character-level errors by 41%
Benchmarked against a random baseline to validate that the model learned decryption structure

Real Estate Market Dashboard

GitHub

Constructed a data pipeline to ingest government statistics, transforming transaction data into time-series data
Segmented the housing market using PCA and K-Means, revealing regional and socio-economic divides without relying on location-based data
Visualized complex datasets, using interactive maps and graphs to enable exploratory analysis of price and returns

Presentation on Algorithmic Information Theory, Kolmogorov Complexity

Slides

Independently studied literature on Kolmogorov complexity and compression-based similarity measures
Implemented a simple illustrative example (Normalized Compression Distance) to support explanation
Presented core theory, motivation, and limitations to a mixed undergraduate-faculty audience

London Stock Exchange Group, Financial Essentials

(Aug 2025)

Completed a certification program covering the foundational principles of financial markets, investment, and their economic and regulatory drivers

Duke of Edinburgh Gold Award

(Oct 2022-Jul 2024)

Adapted quickly to new challenges, learning navigation, and teamwork in unfamiliar environments
Independently planned and completed an expedition, demonstrating resilience, and self-sufficiency