Tom Murray

The code for the Angular file for a status bar component. Typical input for the component is the current job that the app has open. Depending on the status of the job, this will update the status of the component, which in turn changes how far along the status bar is.

A simple trip editing app where you can create, edit, and delete different trips. It was my first app built using React/Redux.

I decided to build a game using the Ethereum blockchain protocol. The game I built was a battleship type of game, mainly due to its turn-based nature.

As the Ethereum blockchain mines the blocks every 14 seconds, a move could only be made once every 14 seconds. I built the front end of the app using Angular. The Truffle framework was used on the front end to interact with the blockchain.

I built a game using the Corona game engine. The game was written in Lua. It is a fast-paced game where you have to swipe to change the direction of the main character. I programmed the physics of the game myself.

For my master's thesis at Imperial College London, I had to build a neural rat that was able to learn the features of a rat's neural activity that mapped to its position in a square box.

It was a combination of an auto-encoder to learn the features and a recurrent net to map these features to the position. The result only had a 10% error when predicting the position.

I built a bot that performed an arbitrage on the EOS ICO. It was implemented using AWS Lambda.

Before each daily window closed, we checked the current money invested during that window. Using linear regression, we determined if we should invest in the ICO during that window. This involved creating, signing and sending a transaction to the EOS ICO smart contract. We then had another Lambda function that sent the EOS to bitfinex, where it was then sold for a profit. During the six months that the bot was in operation, my Ethereum balance increased 250%.

I worked as the tech lead at a startup called Repairly, which was a company that did on-demand tech repair.

I built most of the software, which included a customer-facing interface, an admin dashboard for the inhouse team, a dashboard for the repair centers and an app for the drivers. The components were hosted on Heroku and AWS.

This project consists of a retrieval-augmented generation (RAG)- based chatbot. A user can upload a document and chat with the bot about its details. It was built using LangChain and LanceDB as the vector store.

Designed and implemented an advanced reinforcement learning system for automated options trading strategies using Ray RLlib and PyTorch. The platform leverages deep neural network architectures like long short-term memory (LSTM), GRU, and custom StackGRU to process complex market data and learn profitable trading patterns. Key features include:

• Multiple specialized environments for single-option, contract-type, and portfolio trading strategies
• Sophisticated observation processing with price normalization, volatility modeling, and technical indicators
• Configurable reward functions based on profit metrics with customizable risk parameters
• Comprehensive position management with delta, gamma, theta, and vega thresholds
• Modular architecture allowing for easy strategy experimentation and extension
• Training pipeline with checkpointing, GPU acceleration, and performance monitoring

While not achieving market-beating returns, the project served as a valuable testbed for applying RL to financial markets and understanding the challenges of trading algorithmic options in real-world conditions.

A sophisticated database query assistant built using LangGraph and LangChain that allows users to interact with a PostgreSQL database using natural language. The agent understands database schema, translates natural language queries into SQL, and returns formatted results. The architecture leverages a state graph workflow to orchestrate interactions between the LLM (locally running Llama 3.2 via Ollama), database tools, and user queries.

Key features include database schema introspection, multilingual support, and a well-structured agent workflow with proper error handling.

This project demonstrates advanced patterns for building AI agents with local LLMs, tool integration, and complex orchestration, showcasing the practical application of modern LLM application development techniques in a data analysis context.

A sophisticated database query assistant built using LangGraph and LangChain that allows users to interact with a PostgreSQL database using natural language in multiple languages.

The agent understands database schema, translates natural language queries into SQL, and returns formatted results. The architecture leverages a state graph workflow to orchestrate interactions between the LLM (locally running Llama 3.2 via Ollama), database tools, and user queries.

Key features include database schema introspection, multilingual support, and a well-structured agent workflow with proper error handling.

This project demonstrates advanced patterns for building AI agents with local LLMs, tool integration, and complex orchestration, showcasing the practical application of modern LLM application development techniques in a data analysis context.

Created a course in Udemy to show people how to build a text generator from scratch with PyTorch. I go through all the basics of finding a dataset, creating a tokenizer, training the generator, and then inference. The end result is a streamlit app where the user can generate a trained model by entering some text.

This project is a sophisticated cryptocurrency trading bot that leverages technical analysis indicators to execute automated trades on Coinbase. The system monitors real-time market data through WebSocket connections, processes candlestick patterns, and applies technical indicators like MACD, Bollinger Bands, and TSI to generate trading signals.

The architecture features a modular design with separate components for data collection (WSListener), candlestick processing (CandleHandler), signal generation, and trade execution (Trader). All market data and trading activities are stored in SQLite for analysis and auditing. The bot supports simulation and live trading modes with configurable risk management through take-profit and stop-loss parameters.

Key technical highlights include a multi-threaded operation for real-time data processing, comprehensive logging, containerization with Docker, and a clean command-line interface for configuration.