A Year 13 student in London. I build models, automation pipelines and web applications, including a data-breach detection system for LLM prompts and the platform behind a watch resale business.
At an international IT company I built a system that logs prompts entered into large language models, flags the data-breach risk in them, and summarises them into portals for review.
For a watch resale business I built the platform: a production storefront, and a pipeline that gathers the details of each purchased watch and drafts its Shopify listing.
For my Arkwright Engineering Scholarship I trained an optical character recognition model on more than 700,000 handwritten characters. It reaches 82% accuracy on characters outside the training set, with most remaining error on inherently ambiguous case pairs such as K/k and 0/O.
A personal VPS and a home server running a photo library, media streaming and automated content management behind a reverse proxy. Currently being extended to smart-home automation using on-device LLM and vision inference rather than cloud services.
Commissioned work for a law firm: LLM agent workflows for inbox management and billing, with automated parsing and organisation of legal documents and case data.
Commissioned work: a system that logs prompts entered into language models, flags the ones that risk leaking data, and summarises them into portals for review.
The systems behind a watch resale business: a production storefront, and an automation pipeline that scrapes the details of each purchased watch and drafts its Shopify listing.
Commissioned work for an accountancy firm: a site distributing a client-facing guide, self-hosted on a VPS, with signups feeding an email pipeline. Python scripts structure contact data from several sources into the database behind it.
An automation system that runs a dropshipping operation end to end, from sourcing through to advertising, driving several storefronts at the same time.
A convolutional network trained on the EMNIST ByClass set for my Arkwright Engineering Scholarship. Augmentation, learning-rate scheduling and early stopping gave 82% accuracy on characters outside the training set.
Scrapes breaking financial news and places trades based on it automatically. It was built to measure how quickly an automated system could react to a news event.
Competition code for our FIRST Tech Challenge robot, plus the team website. Autonomous navigation using AprilTag detection and computer vision.
A browser prototype that transmits encrypted messages over audio, extending the GGwave acoustic protocol with an encryption layer.
Two chatbots talking to each other, Monty Hall and other statistical simulations, and a minimax tic-tac-toe that cannot be beaten.
An account of the dataset, the augmentation and scheduling used in training, and where the model still makes mistakes.
Read the full report