{ "/projects/corkboard": { "title": "Corkboard", "content": "

The internet has made one-man intelligence agencies of all of us. Armed with some names and too much free time, digging around and connecting the dots is a relatively frictionless if time-consuming experience. On occasion, you might find something interesting. Yet it is common to run into hard scaling limits when attempting to map the connections between entities within large networks.

\"Report\"

To address this, I developed a pipeline that combines web scraping with LLMs to identify passages of online text that evidence meaningful connection between two entities. The output of the pipeline is then stored within an Obsidian vault.

\"Report\"

At the Autonomy Institute we have put this capacity to work on a range of projects including a collaboration with the world’s largest trade union federation.

Blog Post

Made in collaboration with Lukas Kikuchi

" } , "/projects/dark-kitchen-cartography": { "title": "Dark Kitchen Cartography", "content": "

Dark kitchens are dedicated sites for preparing delivery-only takeaway meals; a model that has been widely adopted across the food service industry throughout the past decade. Companies make use of extensive urban datasets and algorithmic techniques to optimise the placement of these sites in well-connected non-residential urban locations where overhead and operating costs are minimal. A better understanding of the successes and failures of dark kitchens could provide insights relevant to economic planning, risk management, urban design and food distribution.

\"Screenshot

To address the lack of detailed data on the growth and current state of dark kitchens across the UK I developed the first open map of Deliveroo’s UK dark kitchen estate. The data was acquired through scraping takeaway listing data from locations across each of the UK’s several thousand postcode districts.

Featured in: Focus on Labour Exploitation (FLEX)

Blog post

" } , "/projects/detecting-tear-gas-canisters": { "title": "Detecting Tear Gas Canisters with Limited Training Data", "content": "

Human rights investigations often require triaging large volumes of open source data in order to find moments within image, or video that are relevant to a given investigation and warrant further inspection. 37-40mm tear gas canisters are some of the most common munitions used against protesters worldwide, however searching for images containing tear gas usage online manually is laborious and time-consuming.

\"Report\"

Working with several researchers from Forensic Architecture, we focussed on applying object detection models to facilitate discovery and identification of tear gas canisters for human rights monitors. For convolutional neural network (CNN) based object detection to work, a large amount of training data is required. To achieve our objective, we benchmark methods for training object detectors using limited labelled data: we fine-tune different object detection models on the limited labelled data and compare performance to a few shot detector and augmentation strategies using synthetic data. We provide a dataset for evaluating and training tear gas canister detectors and show how such detectors can be deployed for a real world application such as investigating human rights violations. Our experiments show that fine-tuning state of the art detectors perform as well as the few shot detector, and including synthetic data can improve results. Within this research my contributions involved training several state of the art detectors (Faster R-CNN and RetinaNet) for benchmarking, data labelling and pre-processing as well as report writing.

\"Report\"

Published at: Winter Conference for Applications of Computer Vision 2022

Collaboration with Lachlan Kermode, Chris Tegho and Ashwin D’Cruz

" } , "/projects/facial-feedback": { "title": "Facial Feedback", "content": "

The following portraits are generated from dozens of images of Bella Hadid, Rihanna and Kim Kardashian’s doppelgängers. I used dlib (C++ machine learning library) to detect each celebrity’s facial landmarks and used these constellations of coordinates as a focal lens to fine-tune and average the faces of their doppelgängers into a collective identity.

\"Bella

Bella wears Jennifer Lawrence, Carla Bruni, Isabelle Mathers and Joann van den Herik

\"Rihanna

Rihanna wears Priscila Beatrice, Andele Lara, Renee Kujur, Yna Sertalf and @notavrey

\"Kim

Kim wears Kami Osman, Sonia Ali, Amra Olevic Reyes, Claire Leeson, Naya Rivera, Thalia Almodovar, Jelena Karleuša, Jelena Peric, Larsa Pippen and Monyy Monn

This small project arose from my interests in the aesthetic intersections of doppelgängers within science fiction and religion, celebrity culture and social media induced feedback cycles. I am interested in further exploring how platforms mediate monocultures and outliers.

" } , "/projects/fossil-forest": { "title": "Fossil Forest", "content": "

\"Report\"

Single use plastic water bottles are the ghostly artefacts of vast supply chains constructed upon myths of purity. Despite their limited utility, they are grimly charismatic objects and can be found everywhere. There exists myriad designs and each bottle has playful sonic properties.

Inspired by Jugaad robotics, I designed a tree-like structure that reacts to the environmental conditions where bottled water is sourced. Built from bottles, computer fans, servos and sewage piping, each branch represents a location where water is bottled and generates sound responding to rainfall conditions using the OpenWeatherMap API.

Syrup Zone · Fossil Forest

Dry conditions are sonified through the creaking of steel wire across the knurls of a bottle cap echoing through the chamber of an empty bottle. These servo driven noises are an attempt to mimic dry trees creaking. Rainy conditions are sonified through the dull moan of airflow over a slit cut into the side of an empty bottle. With enough time I would build an entire forest of these trees to orchestrate the sounds of bottles from locations around the world.

\"Detail \"Another
" } , "/projects/project-2025-index": { "title": "Project 2025 Index", "content": "

\"Report\"

Project 2025 was a controversial plan developed by the Heritage Foundation think-tank that outlined a policy framework for a future conservative US president. The 900-page publication lacked an index, so I created one to provide researchers with another approach to navigating the contents of the document.

The index was developed as a FastHTML app with a vector database to allow users to go beyond keyword search. Each indexed item contained LLM-generated summaries accessible via tooltip situating each item within the document.

Featured in: Progressive International & DeSmog

Project 2025 Index

Blog Post

" } , "/projects/risks-to-british-business": { "title": "riskstobritishbusiness.today", "content": "

\"Risks

Risks to British Business reads the annual report filings that UK companies lodge with Companies House and the Financial Conduct Authority, and pulls out the impact events buried inside them: the storms, supply-chain shocks, conflicts, regulatory changes and other external forces that businesses describe as having hit their operations. Each event is extracted, verified, traced back to the original filing, and curated by a human analyst.

The platform has scanned 3.6 million filings over the past financial year, surfacing impact events across more than a thousand UK companies, with coverage updated daily in step with Companies House and FCA releases. It processes every electronic filing these regulators publish, giving comprehensive coverage of most UK SMEs today, and every UK company by 2027. Additional context is drawn from company websites and news reporting.

Findings are presented three ways: an interactive map plotting where impacts land across the UK; an impacted products view that traces events down to the specific own-brand goods affected; and a taxonomy that organises everything into a navigable graph of event types, themes and downstream impacts.

While AI does the heavy lifting in the extraction pipeline, every insight published to the dashboard is validated, sourced, and signed off by a human subject-matter expert, holding accuracy at the scale of thousands of documents a day.

View Demo

\"The

\"The

" } , "/projects/scribe": { "title": "Scribe", "content": "

Scribe is an AI-powered research tool that dramatically speeds up the analysis of free-form text, transforming weeks of work into actionable insights in minutes. Unlike generic text analysis tools, Scribe is purpose-built for the think tank and public policy sector, by researchers from the sector.

Qualitative data from surveys, interviews, and focus groups is rich in insight but notoriously hard to analyse. Current methods are slow, manual, and prone to error, often leaving valuable data untapped.

Leveraging advanced language models, Scribe transforms large-scale text responses into actionable data by summarising, categorising, and embedding them for seamless navigation between macro-level topics and micro-level perspectives. I developed the initial prototype for Scribe using data from the world’s largest shorter working time trial co-led by the Autonomy Institute.

Demo

Made in collaboration with Lukas Kikuchi

" } , "/projects/six-degrees-of-reform": { "title": "Six Degrees of Reform", "content": "

\"Report\"

Reform UK brands itself as an “entrepreneurial political start-up” with multi-millionaire real-estate tycoon and party chairman Richard Tice advocating for more business leaders to engage in politics and improve how the country is run. This prompted me to ask:

What British companies have Reform’s leadership and their extended network started-up over the past decade?

To answer this, I conducted an analysis of Companies House data, identifying notable companies established since 2014 in proximity to Reform.

Featured in: The Guardian

Blog Post

" } , "/projects/sugartrail": { "title": "Sugartrail", "content": "

\"Report\"

Sugartrail is an open-source network analysis tool to empower OSINT researchers to navigate corporate ownership networks using data from the UK Companies House register. The tool was developed within a self-directed 2-month tech fellowship awarded by the investigative journalism group Bellingcat and funded by the Alfred Landecker Foundation.

Featured in: Bellingcat, Global Investigative Journalism Network & Journal of Economic Criminology

" } , "/projects/their-capital-at-risk": { "title": "The Rise of AI as a Threat to the S&P 500", "content": "

\"Report\"

Two years after the launch of ChatGPT, AI is used in some capacity by most major companies. Its rapid adoption brings new risks, some of which remain poorly understood.

I authored the following report examining changes to the “Risk Factors” sections of all S&P 500 10-K filings between 2024 and 2025 to understand how AI-related risks are being described. I draw particular attention to subtle revisions in reporting which are easy to miss in dense legal language but often indicate shifting priorities and emergent anxieties.

\"Update

I used textual processing algorithms and large language models (LLMs) to align and compare filings, and created an interactive dataset to explore the results.

\"NumberFeatured in: Stanford Law School, The Information, Cybersecurity Law Report, Investopedia, Ed Zitron’s Where’s Your Ed At newsletter & the Register

ReportDataset

" } , "/projects/twilight-watcher": { "title": "Twilight Watcher", "content": "

\"Report\"

During the 2020 lockdowns, I discovered a website streaming footage from thousands of unsecured security cameras around the world. I was drawn in by the hypnotic quality of these static scenes. Deserted car parks, storefronts with an occasional shop-keeper wearing a mask, empty landscapes. The quietly dramatic changes in light at sunrise and sunset were often the most interesting event broadcasted from at each location.\"Report\"Inspired by the arsthetics of this ‘slow cinema’, I began analysing the transformations in colour from CCTV footage at twilight. Focusing on a weather camera at the Calar Alto Observatory in Spain, I programmed my Raspberry Pi computer to record the landscape each evening. I then applied colour-sorting algorithms to the footage, generating unique patterns that reveal the subtle variations between nights. The results demonstrated how small differences in weather, atmospheric conditions, and digital artifacts could render entirely distinct chromatic signatures.

\"Report\"

" } }