My Work

This article by Flank discusses the evolution of foundation models and the impact of new models on the AI landscape.

This article by Flank discusses the limitations of RAG and the need for a more sophisticated approach to building AI-powered agents.

Good weight initialisation is an important step in successful training of Artificial Neural Networks. Over time a number of improvements have been proposed to this process. In this paper we introduce a novel weight initialisation technique called the Straddled Matrix Initialiser. This initialisation technique is motivated by our assumption that major, global-scale relationships in data are linear with only smaller effects requiring complex non-linearities. Combination of Straddled Matrix and ReLU activation function initialises a Neural Network as a de facto linear model, which we postulate should be a better starting point for optimisation given our assumptions. We test this by training autoencoders on three datasets using Straddled Matrix and seven other state-of-the-art weight initialisation techniques. In all our experiments the Straddeled Matrix Initialiser clearly outperforms all other methods.

This article by Weaviate describes my significant contribution as a lead machine learning engineer at Moonsift, where I implemented a semantic search system using a vector database and chose Weaviate for its efficiency and scalability.

This blog post discusses the implications of OpenAI's new Assistants API and custom GPTs on the AI agent and copilot market, highlighting potential impacts on different types of AI businesses and the overall advancement of conversational AI experiences.

The article focuses on the work I have done at Moonsift to comprehend and map consumers' online shopping patterns using advanced multi-modal AI.

This notebook shows you how to set up a free LLM inference server in a Colab notebook, bypassing previous limitations and complexity. Using Ollama and ngrok, you'll get a public URL for an inference server that supports Orca-mini. Ideal for prototyping but not for production use.

Art Affinity is an AI powered search engine that allows you to search for art using images and text. The search focuses on the actual content of artworks and not keywords or meta tags. I built this using: ChromaDB, React and the OpenCLIP implementation of CLIP.

In this article, we delve into the complexities and challenges of utilising AI for product discovery in a cross-retail setting, discussing the limitations and intricacies of using multi-modal models like CLIP for embedding product data from thousands of online retailers.

This articles explains the paradigm shift in search from keyword-based systems to advanced semantic search powered by AI and machine learning, highlighting the limitations of traditional search and the transformative advantages of semantic search.

This article explores how semantic search has been enabled by vector databases, detailing their benefits in e-commerce and evaluating key contenders on performance, maturity, and ease of use.

This writing charts the evolution of NLP from foundational statistical methods like N-grams to the groundbreaking capabilities of GPT-4, delving into the transformative journey towards advanced language understanding by machines.

The project used Automatic Identification System (AIS) data and LSTM neural networks to quantify and predict congestion at major maritime ports.

We showed that fine-tuning on our more robust dataset and employing an alternative loss function can lead to improved detection of synthetic images generated by Stable Diffusion models.