About

I am Boyu Hou, and most people who know me in English call me Billy. At Delia School of Canada, on the Ontario curriculum, I took Business Leadership: Management Fundamentals, Calculus and Vectors, and Financial Accounting Fundamentals. These were my first courses in finance. However, I did not continue with finance at university. Computing and AI interested me more, so I studied Computer Science and Electronics at Bristol. Later, the models I cared about most turned out to be easiest to test on financial markets. That brought me back to the grounding I had started with.

Houmoon

I founded Houmoon Ltd in April 2024, while I was still at Bristol. The product was a mobile app for everyday reflection, built with React Native and Firebase. In it, a conversational AI guided short reflective sessions alongside journaling. About 100 people used it in internal testing, but it was never released publicly. I closed the company in May 2026 for two reasons. First, the product sat close to the line of UK medical regulation. Second, I had planned to stay in the UK on a start-up visa after graduating, and I decided instead to apply for graduate study in the United States.

From speech enhancement to learning systems

In my final year at Bristol, I worked in a group of six on causal audio-visual speech enhancement with Mamba. I then built VDBC-Mamba-2 on my own. The paper is now submitted to IEEE ICASSP 2027. Speech enhancement was where I started, but the part that held my attention was the model itself. A state-space model carries a hidden state forward as each new frame arrives, so what it knows at any moment depends on everything it has already seen. Therefore, I began to ask what it would take for a deployed system to keep learning from the data it receives.

Now

This is the direction I want to follow in graduate study. I want to build AI that maintains state, revises its beliefs and adapts with low latency as conditions change. Financial markets give a strict test of this. A model there has to act before the data are complete, and each forecast is later checked by the prices that follow. My VaR and HMM projects are a first step, and both were rechecked so that every forecast uses only past data. Furthermore, Living Through History, a planned experiment, would move a learning system through market history from about 1900 onwards without letting it see the future. I do not yet know which architecture is right for this.