Zijun (Brighton) Liu

I am an undergraduate at the University of California, San Diego, pursuing a B.S. in Astrophysics with a minor in Data Science. Before transferring to UCSD, I completed two years at UC Davis double-majoring in Physics and Data Science, graduating with a 3.84 GPA. At UCSD, I carry a 3.98 institutional GPA.

Currently, I am an AI/ML Engineer Intern at XView LLC, where I architect multi-agent LLM pipelines and engineer RAG systems for production grounding. Previously, I was an AI/ML Quantitative Research Intern at Rothenberg Wealth Strategies, building TCN-based probabilistic forecasting models and a three-stage signal pipeline over ~10M rows of market data. Prior to that, I interned at Himiway Intelligent Technology as a Data Science Intern, deploying XGBoost demand-forecasting models that accelerated executive decision cycles from four weeks to one week.

My background in astrophysics shaped one skill above all others: extracting weak signals from overwhelming noise. Processing million-row galaxy spectra trains a particular kind of skepticism — every pattern is a hypothesis until proven otherwise, out of sample. That same discipline carries into my ML and quant work: at Rothenberg, I built an 8-layer TCN for probabilistic quantile forecasting alongside a three-stage signal pipeline combining a Conv1D-VAE with a meta-learning gatekeeper over ~10M rows of US equity data.

As a 4D Go player and the lead developer of the KataGo × LLM project, I think about both ML systems and quantitative research through the lens of variance management. Modern Go AI teaches a profound lesson: optimize for the probability of winning, not the margin of victory — it actively sacrifices high-margin local gains to reduce global variance. I apply this same philosophy to model and strategy design, favoring lower absolute returns with minimal drawdown and high information ratios over brittle high-return configurations. In both Go and production ML, managing variance to survive to the next iteration is the ultimate edge.

I am also drawn to world models as a research area — the question of how an agent can compress its environment into an internal representation good enough that decision-making becomes almost trivial. What I find most compelling is how much the right inductive bias matters: a Transformer trained naively on planetary trajectories learns to curve-fit the ellipse, but restricting its attention to recent timesteps — encoding the Markov structure of Newtonian mechanics directly into the architecture — forces the model to learn the actual causal chain of force, acceleration, and velocity change. My physics background makes these architectural choices feel intuitive rather than arbitrary. I have been following the research trajectory from Ha & Schmidhuber’s 2018 World Models paper through PlaNet and Dreamer, and I occasionally write about these ideas on LinkedIn (post 1, post 2).

If you are looking for an engineer who builds robust ML systems while thinking carefully about what the models actually prove, get in touch.