A Llama-style language model written from scratch in PyTorch, small enough to pre-train on a laptop, then fine-tuned with reinforcement learning.
Appendix A. All projects
Everything I have built that I am happy to discuss, grouped by topic. A selection, with figures, is on the main page under §1 Selected work.
LLMs and retrieval
A local assistant that answers programming questions from a code corpus, reviews its own answers and retries when they fall short.
Local RAG chatbot
Chat with your own PDFs, fully offline. LangChain and FAISS handle retrieval, a local Ollama model (Phi-3) writes the answers, and Streamlit provides the interface.
Reinforcement learning
A reimplementation of the MOReL offline reinforcement learning framework with the improvements I proposed in my MSc thesis. It outperforms the original on the D4RL HalfCheetah benchmark.
Neural-network ensembles that learn how an environment evolves from offline data, comparing deterministic and probabilistic models. They are the learned models behind J-MOReL and MBiCEM.
A trajectory planner that runs iCEM on learned dynamics models and heavily penalises the predictions the model is unsure about.
Computer vision
Detects sponsor logos in F1 broadcast footage with YOLO, as a first step towards measuring each brand's screen time. Zero-shot YOLO-World finds no real logos; a YOLO11s fine-tuned on 69 labelled frames reaches mAP50 0.61 on held-out frames.
Data and MLOps
An end-to-end machine learning system, from Spark data processing to a public forecasting app on Google Cloud.
Tools and other work
Matplotlib GitHub theme
Light and dark Matplotlib themes that match GitHub's colours, so plots in a README follow the reader's theme. Most of the plots on this site were made with it.
View My Satellites
A Unity app that shows where Earth-orbiting satellites are right now, in an interactive 3D scene. Built by a team of students during the NASA Space Apps Challenge.