VDBC-Mamba-2
Visual-Delta Band Conditioning for Causal Audio-Visual Speech Enhancement
The visual term of the conditioning module is exactly zero when the visual input is zero. Robustness is tested with black, shuffled and reversed video.
Resilience matters. Nothing I have lived through is ever truly left behind. It becomes part of who I am, and I grow with it. The future is better not because the past disappears, but because I carry it forward and become more than I was.
Boyu Hou is a BEng Computer Science and Electronics graduate of the University of Bristol (2022–2026). The research focus is placed on AI systems that keep learning after deployment and make low-latency decisions as new information arrives. Financial markets are chosen as the test. Information there arrives in a fixed order, and each decision can be scored against what happened next. A single-author preprint on causal audio-visual speech enhancement, VDBC-Mamba-2, has been submitted to IEEE ICASSP 2027, and two projects on market regimes and Value-at-Risk have been completed. In addition, Houmoon Ltd, an AI wellbeing start-up, was founded in 2024 and closed in 2026 before any public release.

2026
Single-author preprint on causal audio-visual speech enhancement, submitted to IEEE ICASSP 2027.
2025–2026
Regime-aware VaR and HMM regime detection on market data, backtested using only past information.
2024–2026

Founder of an AI wellbeing start-up. About 100 internal testers; dissolved in May 2026 before any public release.
2022–2026

BEng Computer Science and Electronics
Final-year group project on causal audio-visual speech enhancement with Mamba.
2019–2020
International Foundation Year in mathematics, computing and academic English; top 5% of the cohort.
2017–2019
Ontario curriculum, including Calculus and Vectors and Financial Accounting Fundamentals.
Visual-Delta Band Conditioning for Causal Audio-Visual Speech Enhancement
The visual term of the conditioning module is exactly zero when the visual input is zero. Robustness is tested with black, shuffled and reversed video.
Four one-day VaR forecasts for SPY, backtested on 2023–2025 data using only past information.
Market regimes from a Gaussian HMM with forward filtering, tested as state-conditional VaR on ten assets.
Six-person Bristol coursework on causal audio-visual speech enhancement with a Mamba backbone.
I thought I was looking at glass. But the observation itself did not prove that it was glass.
A planned experiment in learning from a historical timeline without seeing the future.