Mini Projects Portfolio

This is a repo I used to learn new tools and try out project ideas. Below are some highlights:

AI and Agents

MilkBot

This is an AI shopping assistant for an online milk store. It can answer questions about milk products, place items into and out of the shopping cart, and initiate checkout.

Tools:

• The catalog of milk products was scraped from the Woolworths website. I had to use headless browsing with Puppeteer because the website loads dynamically.
• The product information was embedded using the OpenAI API with the text-embedding-3-small model, and stored in a ChromaDB vector database.
• The AI agent was implemented with LangChain.
  • Retrieval-Augmented Generation (RAG) capabilities were implemented where information about products relevant to the user's question was retrieved from the ChromaDB database and augmented to the user's question.
  • Adding and removing items from the shopping cart were implemented as agent tools.
• The backend was implemented with Flask.
• The frontend was implemented with React.
• The LLM is GPT-5-nano via the OpenAI API.

PeanutBot

GPT-2 finetuned to replace alphanumeric characters with peanut emojis (🥜).

Example conversation:

👋: Hello, how are you?
🤖: 🥜🥜! 🥜🥜🥜 🥜🥜.
👋: Who was the first president of the United States?
🤖: 🥜🥜🥜 🥜🥜🥜🥜 "🥜🥜🥜 🥜. 🥜🥜🥜🥜". 🥜🥜🥜🥜 🥜🥜🥜🥜 🥜🥜🥜.

Tools:

• Finetuning using Python and Hugging Face.
• Backend implementation with Flask.
• Frontend implementation with React.
• Agent implementation with LangChain.

Simpsons RAG

A chatbot that steers conversations toward Simpsons episodes while denying it's doing that.

• Scraped Simpsons episode data from IMDb.
• Generated semantic embeddings of each episode's description using OpenAI's text-embedding-3-small, and stored in a ChromaDB database.
• The RAG was implemented with LangChain.
• The frontend was implemented with React.
• The backend was implemented with Flask.
• The LLM is GPT-5-nano via the OpenAI API.

Data Analysis and Data Visualisation

US Houehold Income Data Visualisation

I was contracted to research and develop an animated data visualisation of US household income data and recreate/update a classic Financial Times animation.

Tools:

• Data wrangling and exploration with pandas, matplotlib, and seaborn.
• The developlment process is documented in a Jupyter Notebook
• A frontend for constructing customisable animated data visualisations with html, css, and javascript.

I want to emphasise that this wasn't a simple matter of plotting a csv. Here are some of the issues that needed navigating:

Figuring out what exactly the variables represent
Inflation, which requires CPI-based adjustments
Some incomes are replaced with codes, such as "999999999" meaning "this individual is under 15 and outside the scope of the survey." Different variables have different codes.
Making sure households aren't double-counted from multiple individuals in the same household
High income values are obfuscated for privacy, but this has been done differently in different years. In early surveys incomes were simply truncated above some value, but since the 90s there's been replacement values, and rank proximity swapping and I haven't spent much time figuring out how that works.
Samples need to be weighted because some households are more likely to be surveyed than others
There is both cross-sectional data and longitudinal data
There are inconsistencies in how surveys were conducted over the years.

3D Model Size Visualization

This is a pretty basic 3D visualisation frontend for visualising the relative sizes of notable AI models (in terms of parameter counts) in three dimensions using Three.js.

Tools:

• The website uses Three.js for the 3D visualisation.
• The data is retrieved from the Epoch AI database.