cv

B.S. Astrophysics, Data Science Minor

• GPA: 3.92 (Institutional) | 3.87 (Cumulative) · Phi Beta Kappa
• Applying to M.S. programs in CS / ML for Fall 2027; available for a Summer 2027 internship, Jun - Sep 2027.

Physics (Astrophysics Emphasis), transferred to UCSD

• GPA: 3.84

What does self-supervision add to engineered features for stellar light curves?

• Accepted at the ML4PS workshop @ NeurIPS 2026. arXiv · Project · Code

Machine Learning Research Intern

• Deployed a LightGBM model I trained and selected on walk-forward validation as an unattended production service on a cloud VM: scheduled data ingestion, feature-build and inference phases, duplicate-safe external API calls with daily reconciliation, and a FastAPI + DuckDB ops dashboard.
• Made the live system compute features exactly as in training: fixed feature order and automatic rejection of missing or non-finite inputs.

AI Engineer Intern

• Built a 5-agent CrewAI content pipeline with retrieval over product docs (Chroma), generating 35 platform-tailored drafts per run across 5 social platforms.
• Set up the evaluation loop (prompt A/B tests, chunking-strategy comparison) that cut factual errors by about 60% and brought the human revision rate of generated drafts to about 30%.

Machine Learning Research Intern

• Ran controlled LightGBM experiments over ~2.4M rows of time-series data: feature engineering and ablation across 1,000+ features, label and loss-function variants including a custom objective, and hyperparameter sweeps (Latin hypercube and Optuna) evaluated by walk-forward validation.
• Built the experiment tooling in Python: hyperparameter sweeps defined from config files, a multi-worker queue runner (1,000+ runs per sweep), per-run loss-curve diagnostic notebooks, and an HTML report builder.
• Built a self-supervised VAE with a temporal-convolutional encoder and decoder over ~10M rows of time-series windows.

Data Science Intern

• Refactored notebook analyses into a production ETL and forecasting codebase (SQL/Python, argparse, logging, Makefile) for continuous ingestion.
• Trained and deployed an XGBoost model for store-location and inventory decisions, cutting the decision cycle from 4 weeks to 1.
• Designed and evaluated a months-long in-store A/B test (fat-tire vs. thin-tire e-bikes), analyzing conversion rate and revenue per visitor in Tableau.

Self-Supervised Learning on Stellar Light Curves

• First author, with Y. Ma and Prof. C. Theissen (UCSD). Paper accepted at ML4PS @ NeurIPS 2026.
• Pre-trained a convolutional encoder with a GRU latent-prediction objective (PyTorch) on unlabelled TESS light curves; froze it and evaluated the embeddings with linear probes on 10 downstream tasks.
• Concatenated the embeddings with 25 engineered features, improving linear-probe results on 7 of 10 tasks and beating a Conv1D and an MLP trained end-to-end; multi-seed experiments with bootstrap confidence intervals, tracked with Weights & Biases.

KataGo × LLM — LLM Post-Training & Edge Deployment

• Project lead of a volunteer team. Post-trained Qwen3-8B (LoRA, Hugging Face TRL, vLLM rollouts, Dockerized training) to play 9x9 Go and explain its moves by distilling the KataGo engine: GRPO first, then a rules-SFT / play-SFT / GRPO curriculum on 302k engine-verified examples; first wins vs a classical engine, zero illegal moves.
• Built the evaluation (board-perception probe, top-k agreement, refereed engine ladder) and stabilized GRPO; removed reward hacking, countered entropy collapse with DAPO clip-higher + adaptive entropy control.
• Deployed 4-bit GGUF (llama.cpp) on an 8 GB-VRAM GPU behind a C++ GTP proxy for the Lizzie GUI (~42.5 tok/s, 0.52 s TTFT); served base + 2 LoRA adapters from one vLLM instance to run 3 bots live on OGS. Demo

ASTR199 — Deep-Learning Inference of Stellar Parameters

• With Prof. C. Theissen, UCSD.
• Engineered a DuckDB ETL pipeline cross-matching ~1M rows across 7 photometric catalogs; built a physics-informed feature space of color indices and absolute magnitudes.
• Built a two-stage deep-learning pipeline that resolved multi-task training collapse with a sequential design and physics-informed features (log g R² 0.52 → 0.83).

Breaking Barriers Hackathon (AWS × Deloitte × AT&T) — Finalist, Top 8 of 32

• Built a real-time anomaly-detection pipeline: sliding-window spatio-temporal features and an XGBoost classifier with thresholds tuned on PR-AUC for imbalanced data.
• Served inference on AWS Lambda (under 400 ms), streaming alerts to an S3-hosted map dashboard.

PHY199 — Galaxy Spectra PCA Pipeline

• With Prof. Wittman, UC Davis.
• Automated Python/R preprocessing pipeline for 2,300+ high-dimensional spectra; PCA-based outlier filtering with scikit-learn, interpreting PC1–PC3.

• Phi Beta Kappa, UC San Diego chapter

• Breaking Barriers Hackathon Finalist — Top 8 of 32 Teams (AWS × Deloitte × AT&T)

Project Lead

President & Co-Founder

• Grew membership to 30+ members; coordinated with ACGA and universities to run inter-school competitions and community events.
• Led cross-cultural, mixed-skill communications; organized weekly programs and external collaborations.