All Projects

Flagship project and CV centrepiece. A fully automated, end-to-end data ecosystem — not just a chatbot. The system monitors the HM Land Registry data source, runs automated data engineering pipelines, manages and updates its own database, performs statistical analysis across 1.76M transactions and 16 years of London property data, and delivers answers directly to users across all 33 boroughs. Every layer — from raw data ingestion to the user-facing response — is self-sustaining and runs without human intervention.

Automated monthly data pipeline publishing updated Kaggle datasets and mini-reports from HM Land Registry Price Paid data. Runs every 5th of the month.

Built a personal automation app using OpenClaw on macOS — AI-powered inbox management.

Optimised Google's Gemma-3-270M via structured pruning and knowledge distillation — exploring model compression techniques for efficient LLM deployment.

Fine-tuned GPT-OSS 20B on PubMedQA with chain-of-thought reasoning, achieving 73.6% accuracy — rank 19th on the PubMedQA leaderboard. Trained via QLoRA on RTX 4090.

Tutorial on building a composable AI agent using the Model Context Protocol (MCP) for stock investment analysis.

Hands-on guide fine-tuning OpenAI's GPT-OSS 20B model for multilingual reasoning on an RTX 4090.

Fine-tuned GPT-OSS 20B into a distinct debate personality via QLoRA on RTX 4090. Full pipeline: dataset curation, fine-tuning, Transformers deployment.

AI-powered registration system for local government. FastAPI backend deployed on Render, Gradio frontend on HuggingFace.

End-to-end series covering motivations, machine setup on WSL2, and results from fine-tuning open-source LLMs for Text-to-SQL tasks.

Kaggle competition. Mathematical reasoning with LLMs at olympiad level. Highest personal ranking.

Awarded the AI Skills Challenge badge from Microsoft Azure.

Completed the "Operationalizing LLMs on Azure" certificate course by Duke University.

Earned the Fundamentals of Azure AI Services badge in the AI Skills Challenge.

Real-world use-case article and code: building an AI-powered research assistant with a modern RAG stack.

Self-initiated pilot exploring a local LLM for council email handling, presented as an AI project proposal to Wandsworth Borough.

Article dissecting the mathematics of positional embeddings in NLP models. Uses a Crayola colour-blending analogy to show how embeddings represent words as numbers. Independently worked out how the positional encoding matrix is structured: each row is a word position (k), each column is a dimension (i), computed via alternating sine and cosine — then added element-wise to the semantic embedding. Published on Medium.

Conversational chatbot capable of recalling previous parts of the conversation, such as a user's name.

LLM-powered tool for generating sales copy, tweets, and product descriptions for different target audiences. Built as a solo project, PitchPal held its own in a direct face-off against Attentive®, a funded marketing AI — a moment that proved the work was on the right track.

Chatbot with the personality and vocabulary of a 5-year-old, built with LangChain and OpenAI.

Designed and trained a small language model end-to-end, sharing the process and insights with the community.

In-depth Kaggle tutorial on 4-bit quantisation and its impact on Llama2 performance.

Kaggle competition. Multiple-choice science exam solved by LLM. Top 42% out of 4,700+ teams.

Kaggle competition. LLM-based text quality evaluation. Top 32% out of 4,700+ teams.

Topic modelling and collaborative filtering system for action game recommendations.

Social listening project using NLP to analyse sentiment around McDonald's in London.

NLP sentiment analysis model applied to New York restaurant reviews.

Traditional NLP sentiment model trained on L.A. restaurant reviews.

Composed a suite of introductory NLP tutorials covering tokenisation, stemming, lemmatisation, transformers, and attention mechanisms. Published in Towards AI and Towards Data Science.

Built and delivered a BERT-based LLM classification model for the team during my tenure at Microsoft — the first hands-on dive into state-of-the-art NLP. A rewarding year.

Completed Harvard's foundational AI course covering search algorithms, classification, optimisation, machine learning, neural networks, natural language processing, and large language models — all implemented in Python.

Volunteered to build a self-improving sentiment classification web app for Matters, a Hong Kong publisher. The model updated with each user interaction.