Roles, builds, and the toolbox behind them.

fashion-web ↗

Predicts which fashion micro-trends will peak 30 days out. Five live sources (Google Trends, Bluesky firehose, Pinterest, Reddit, YouTube) merge into a Holt-smoothed composite signal tracking 250+ terms across silhouettes, colours, and aesthetics. A logarithmic prediction market (LMSR, b=100) lets traders compete against the house model with virtual StyleCoins. Gemini 2.5 Flash breaks down outfit photos and matches them against runway collections from eight major houses. House predictions are Brier-scored.

Taylor Swift recommendation engine ↗

Six engines running in parallel — Sentence-BERT on lyrics, a variational autoencoder compressing Spotify audio features to 16 dimensions, graph node2vec embeddings, neural collaborative filtering, contrastive self-supervised learning, and a weighted ensemble with a consensus boost for cross-engine agreement. Runs over Taylor's full discography plus hand-curated editorial bridges to related artists. Rebuilt from an R/Shiny original in TypeScript + FastAPI; recommendations in under 200ms.

MFoCS thesis · immune biomarkers

Two problems in automated biomarker detection on multiplex immunofluorescence data from real tumour samples. First: four automated thresholding methods (Otsu, IsoData, custom GMM, minimum cross-entropy) competing to replicate a pathologist's gate — CTLA-4 resisted all of them, because the biology is genuinely ambiguous. Second: spectral clustering with biological co-expression constraints encoded directly in the similarity matrix, beating five of seven standard methods on plausibility. Thesis available in Oxford repository.

JAYA protein feature selection ↗

Modified a parameter-free stochastic optimizer with a size-penalty fitness function and elitism to improve feature selection for protein function identification. Tested on three biological datasets; consistently produced smaller feature sets without sacrificing classification accuracy. Poster PP-28 at INBIX'22.