I train models, build the systems around them, and design how people meet them.

Creative technologist who has just completed an MSc in Applied Machine Learning at UAL's Creative Computing Institute. I build interactive systems end to end: training and fine-tuning the models, engineering the pipelines around them, and designing the interfaces and installations people actually use. Computer vision, generative image models and language models, built into things people can play, draw with and step inside.

Fig. 1 The author, in ten layers of ASCII

Type: MSc thesis, 2026
Field: Neural cellular automata, weight-space interpolation
Built with: PyTorch, Next.js, WebGL
Live: new-natures.vercel.app

Fig. 2 Nautilus: a pattern grown by nature

New Natures

A comparative study of three ways to combine neural cellular automata trained on natural textures, testing whether linear mode connectivity, a property studied on image classifiers, holds for models that grow their output.

The combinations that held became a series of pattern designs, and the best of them run live in an interactive exhibition in the browser.

Read the case study Open the exhibition

Coherent blends by combination method: both parents stay recognisable and the blend never collapses to noise

5 / 10Method 1 · one model trained on both textures

13 / 20Method 2 · two models fine-tuned from a shared start

0 / 10Method 3 · two models trained separately, predicted to fail

5natural textures learned

II.

Type: MSc project, 2026, graded A
Field: Computer vision, projection mapping, live performance
Built with: Python, OpenCV, python-chess, Stockfish

The Rematch

A real chess board that plays back. An overhead camera reads my moves, Stockfish answers at 2750 Elo, Deep Blue's estimated strength in 1997, and its pieces appear on the board as projected light.

I performed it live as a restaging of Deep Blue against Kasparov. The system ran from first move to checkmate in front of an audience without interruption.

Read the case study

Fig. 3 The king: Deep Blue v Kasparov, restaged

7…h6

Deep Blue v Kasparov, 1997, Game 6. Kasparov's seventh move is widely seen as the error that lost the match. The Rematch starts from the position just before it, with me in his seat, and plays it differently.

7Python modules, camera to projector

2750Elo, matched to Deep Blue

0sensors in the board: camera and projector only

1full game played live, start to checkmate

III.

Type: MSc group project, 2026, graded A
My role: Machine learning lead: dataset, LoRA training, ControlNet, inference backend
Built with: Stable Diffusion 1.5, LoRA, ControlNet, rembg, Flask, Three.js

Fig. 4 Mushrooms, a sketch prompt

Borderless

Visitors sketch a mushroom, a planet or a cloud. A fine-tuned diffusion model repaints it in watercolour and places it in a shared 360° world that you can walk into in VR.

I owned the machine learning: a cleaned dataset of around 600 images, four LoRA fine-tunes on Stable Diffusion 1.5, each fixing a named failure in the last, ControlNet sketch conditioning tuned per category, and the backend that turns a browser drawing into a placed asset.

Read the case study

~600training images, cleaned and captioned

4LoRA versions, each fixing the last

6sketch categories

1shared 360° world, explorable in VR

IV.

Type: MSc module project, graded A. My first coding project
Field: Conversational AI, retrieval, character design
Built with: Python, SmolLM-135M-Instruct

Piblo the Plutonian

A Plutonian who makes a 10,000-record space-mission dataset worth talking to. Piblo remembers what you tell it, writes you a story with a language model small enough to run on a laptop without a GPU, and answers questions about real missions in plain English.

Twenty people tested it. Almost all used it without help and called it fun; fewer noticed it was teaching them, which set the brief for the next version.

Read the case study

Fig. 5 Piblo's ride home

PibloGreetings human! I am Piblo from the planet Pluto. What do you call yourself?

SamSam

Piblo…And what do you like to eat? I'm fascinated by what I've heard about earth cuisine.

Samburgers

PibloAlright Sam, I must shoot off back to Pluto! Next time I visit you must make me some burgers to try.

Appendix

Fig. 6 Skull study

From Hand to Algorithm

One of my oil paintings, directed through Midjourney V7 into two product concepts, a chrome candle holder and a ceramic vase. The craft is in the control: balancing image weight and style references so the painting's anatomy and brushwork survive the translation.

See the process

About

I came to machine learning from a First in Social Sciences at UCL, so I start with the person using the system and work back to the model. I'm most useful where engineering, design and research overlap: prototyping the model, building the pipeline, and designing what people actually touch.

Machine learning

PyTorch, neural cellular automata, Stable Diffusion, LoRA fine-tuning, ControlNet, ComfyUI, local LLMs, OpenCV

Engineering

Python, TypeScript, JavaScript, Next.js, React Native, WebGL, Supabase, projection mapping

Design

Interaction and interface design, typography and layout, prototyping in Figma, illustration in Procreate, generative image direction (Midjourney)