Portfolio

Researchers using the SAIL databank often need to match the same person across several health datasets, and many of them don't program. I built a Streamlit app on top of Splink so they can run the linkage and read the results themselves.

Linkage quality went up by 15 to 20 percent. Both this app and the chatbot below are documented in a public GitHub repo so other researchers can reuse or extend them.

Splink is hard to pick up cold, so I built a chatbot that answers questions about the library, linkage theory and the app. It runs Qwen 2.5 behind a Gradio interface hosted on Hugging Face.

I wrote the whole pipeline myself: chunking, embeddings, retrieval and generation. It only answers from the source material, so when the docs don't cover something it says so instead of guessing.

Telling every recent partner you've tested positive is awkward, so it often doesn't happen and the infection keeps moving. I built the tracker at the centre of Glo to do it anonymously.

When someone reports a positive test, it walks their contact graph two steps out and sends each person a note recommending a test, with nothing that reveals who reported. Timing follows STI incubation windows rather than one generic alert for everyone.

I took it from nothing to a working MVP with 100 real users. The team then raised a first round from academic investors at UCL.

Paste up to ten job links and it researches each company, then writes a tailored CV and cover letter straight into a formatted Google Doc. It uses a local open-weights model through Ollama, so there are no API bills and nothing leaves my machine.

Most of the work was making it dependable. I fixed a model that quietly returned near-empty text when the prompt outgrew its context window, a cover letter that leaked into the CV, and a Streamlit page that lost its progress when the tab went idle. I also wrote my own formatter so headings, bullets and links land correctly in the Doc.

For my MSc dissertation I worked inside the Genomics England secure research environment. The question was how mutations in ACMG genes change outcomes for people with cancer and for people with rare diseases.

I built the cohorts, mapped how ICD-10 codes differ in direction and intensity between groups using NetworkX, and tested those differences statistically. As far as I could find, nobody had studied this before.

I worked across several NIHR-funded projects: reports, R visualisations of health gaps between areas and groups, and bid documents that helped the directors secure new funding.

On the CAPITAL bid I ran workshops and focus groups to gather evidence first-hand. I also helped with an entry for the Wellcome Trust's mental health data prize, and built an open-source analytics platform that gives mental health practitioners council-level data.

At an oncology AI company I moved a clinical data pipeline into BigQuery, which doubled data access and query efficiency. I also built a benchmarking framework for the company's LLM-based product that raised quality-control accuracy by 25 percent.

For clinical trials I built a decision-tree framework that sorts breast cancer inclusion criteria by stage. The rest of the job was market research, dashboards and go-to-market planning.