Matt Barty

Presentation on the state-of-the-art generative models and how GenAI is enabling new forms of creative expression. A collection of personal and professional projects that I have worked on.

Procedurally generate beautiful 3D hexagonal landscapes with realistic terrain generation. Runs in the browser with ThreeJS and React.

Self-study implementation of a Denoising Diffusion Probalistic Model (DDPM) for generating MNIST handwritten digits. This implementation was created as a self-study exercise to understand diffusion models through hands-on practice

Developed 3D sandbox game enabling dynamic speech interaction with AI agents in a 24-hour timeframe. Built custom game engine in ThreeJS to facilitate unique AI autonomy and voice-driven gameplay.

Created AI-powered career exploration platform combining interactive maps with LLM-driven country insights within 24-hour hackathon constraint. Implemented Llama-2.1 model through Groq for fast, contextual career advice based on location selection.

Reimagined the classic "Guess Who" game using Speechmatics' conversational AI Flow API. Built interactive web interface with real-time speech recognition and character identification. Engineered dynamic game logic combining speech recognition with AI character reasoning.

One of the first fully AI-directed music videos created just days after public release of txt2vid technologies. Produced with 100% AI-generated visuals and music using cutting-edge generative models, with techniques developed for maintaining character consistency between scenes.

Uses Convolutional Neural Networks (CNNs) to analyze MRI scans and classify them according to tumor type. Achieves strong predictive classification (validation accuracy: 63%, test accuracy: 69%, F1-Score: 0.857) using TensorFlow, Keras, and a pre-trained EfficientNetB0 architecture with advanced image preprocessing.

Established proof-of-concept for AI-assisted product design that improved efficiency and creativity using local Stable Diffusion and Photoshop. Led campaign gathering designer feedback on AI integration and produced marketing content demonstrating AI-augmented design capabilities.

We hack some electronics and cutting-edge multi-modal models to teach my Roomba into avoiding cables.

Trained OCR model on MNIST dataset to recognize handwritten digits in Sudoku puzzles. Features real-time grid detection with perspective transformation, efficient backtracking solver with threading, and AR interface that overlays solutions onto the original puzzle in live camera view.

LLM-enabled Tarot webapp 🃏 Assets generated with Midjourney, interface designed and protoyped in Figma.

A webapp that uses GPT-3.5 and GoDaddy API to generate clever domain names based on your idea description.

Augmenting WhatsApp web with language features, specfically an auto-translate feature with scope to add other features in the future.

Generate bespoke, fully-voiced meditations using GPT-4o and AWS Polly

My usability and clinical data science work on the next generation of surgical robotics.

Generate the perfect user personas for your project using GPT-4o

WebApp that uses BLIP-2 and GPT-3.5 to generate parody LinkedIn posts.

Using python, some complex algorithms, and a webdriver to hack a real-world, online Wordle game.

Combining Midjourney, Stable Diffusion, and Photoshop GenAI models to rapidly create digital art.

A walkthrough tutorial demonstrating how to scrape data from IMDB.com. (ep. 2 of 2)

A walkthrough tutorial demonstrating how to scrape series data from IMDB.com (ep. 1 of 2).

Visualising Pokemon data using sprite art.

Training a DNN to diagnose my own retina scan.

A demo that combines cutting-edge LLM, Image Diffusion, and text-to-voice models to tell an entirely original horror story.

Using Stable Diffusion to create portraits of myself.