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.

My background in astrophysics shaped one skill above all others: extracting weak signals from overwhelming noise. The same rigor I apply to processing million-row galaxy spectra directly informs how I approach financial and ML datasets — with extreme skepticism. At Rothenberg, I conducted falsification on 40 LLM-generated alpha factors sourced from recent academic literature. By implementing out-of-sample vectorized backtesting with realistic transaction costs, I proved all 40 signals yielded negative returns after execution — preventing deployment of multiple spurious strategies. This experience reinforced a core belief: every hypothesis, regardless of its source or apparent authority, demands rigorous out-of-sample stress testing before deployment.

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.

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