Hello there, I'm
Data, AI & Machine Learning
I build machine learning systems end to end, from data pipelines and model training to APIs and deployment.
I'm a developer focused on machine learning and data-driven software. My recent work spans retrieval-augmented generation, NLP classification, LLM security research, and predictive modelling, usually taken all the way from raw data to a deployed, containerised service. I'm currently looking for data science, software and AI/ML engineering opportunities.
Locally hosted retrieval-augmented generation system for querying research papers through a chat interface. PDFs are converted to markdown, chunked, and embedded in a vector database; answers are grounded with per-page source citations. Runs fully offline with local models and Docker Compose.
Predicts a powerlifter's next competition total from 150,000+ IPF records in the OpenPowerlifting database. An XGBoost model cuts prediction error by 17% versus a persistence baseline (19.1 kg MAE, R² 0.959), served through both a REST API and a Streamlit UI.
Detects prompt-injection attacks against open-source LLMs by classifying statistical features of attention weights (entropy, variance, TF-IDF). The best configuration (LLaMA 3.2 1B + Support Vector Machine) reaches 71% accuracy on unseen data, outperforming competing detection tools.
Places news articles on the political compass using a two-stage transformer pipeline: one model first filters for political content, then a fine-tuned DeBERTa classifier scores the article's political leaning, a task general-purpose MNLI classifiers handle poorly.
Cleans the UK financial sanctions list into an analysis-ready dataset: standardising names and aliases, consolidating scattered address fields, and preserving partial dates across ~4,000 individual records, prioritising data integrity over aggressive automation.
I'm open to software and machine learning roles — feel free to reach out or look through my work.