Zijun (Brighton) Liu

I’m Brighton, a senior at UC San Diego studying Astrophysics with a minor in Data Science. I train and ship models end to end: self-supervised sequence models, LLM post-training, and production ML services.

Seeking a Summer 2027 ML internship — sequence models · LLM post-training · production ML.
Applying to M.S. programs in CS / ML for Fall 2027; available Jun – Sep 2027.

Resume — ML Engineering Resume — Research

Highlights

• First-author paper, ML4PS workshop @ NeurIPS 2026 — a self-supervised encoder pre-trained on unlabelled TESS light curves adds to engineered features on 7 of 10 tasks. arXiv · Project · Code
• KataGo × LLM (project lead) — post-trained Qwen3-8B with SFT → GRPO on 302k engine-verified examples; runs 4-bit on an 8 GB GPU and serves base + 2 LoRA adapters from one vLLM instance. Demo · Code
• ML Research Intern, Rothenberg Wealth Strategies — trained a LightGBM model and deployed it as an unattended production service; built a multi-worker experiment runner (1,000+ runs per sweep).

Selected projects

Self-Supervised Learning on Stellar Light Curves

First-author paper, ML4PS @ NeurIPS 2026. A self-supervised encoder pre-trained on unlabelled TESS light curves adds to engineered features on 7 of 10 tasks.

KataGo × LLM — LLM Post-Training & Edge Deployment

Post-trained Qwen3-8B (SFT → GRPO) to play 9×9 Go and explain its moves; 4-bit edge deployment and multi-LoRA vLLM serving.

Research interests

• Self-supervised representations of sequences — what an encoder trained without labels adds to the features domain experts already use.
• World models and inductive bias — how architecture choices (e.g. restricting attention to recent timesteps) push a model to learn the dynamics rather than curve-fit them. I occasionally write about this on LinkedIn (post 1, post 2).
• LLM post-training and evaluation — reward design, reward hacking, and evaluations that show what a model actually learned.

Questions or opportunities: get in touch.