Education
September 2024 - June 2028 · Los Angeles, CA
Work Experience
Research Assistant · Prof. Kai-Wei Chang, Nanyun Peng, and Terence Tao
Built a long-horizon multi-agent reasoning harness for autonomous mathematical research, orchestrating specialized agents with statically enforced tools and capabilities. Designed persistent graph-based memory and iterative planning to preserve competing proof directions and prioritize high-value branches.
Research Assistant · Natural Language Processing
Built a privacy-preserving safety guard for LLM multi-agent systems using an edge-featured graph attention network trained through federated learning without sharing raw interaction traces. The 0.5M-parameter, approximately 50 ms CPU sidecar outperformed the centralized-training ceiling on three benchmarks and reduced live attack success by 43% at above-unguarded utility. Amortized MAS topology design into a 16-entry VQ-VAE codebook with a reward-weighted code predictor and an execution-grounded MLP reranker, achieving the best accuracy across all six benchmarks (84.6 vs. 83.0), over 120x faster design at 2.4 ms, and 21.9-33.2% fewer LLM tokens; both first-author papers are in submission to AAAI 2027.
Research Assistant · AI and Data Mining
Developed Gaussian speculative decoding in PyTorch using lightweight draft models and single-pass parallel verification, accelerating pretrained time-series foundation models by up to 3.1x across five model families without retraining. Proved throughput and forecast-fidelity guarantees; first-author paper submitted to AAAI 2027.
Full-Stack Developer Intern
Built grade management, course management, and role-based notification modules for a university web platform using Vue, Java, Spring Boot, MySQL, and Spring Data JPA.
Data Analysis Intern · Wealth Management
Analyzed 2,000+ high-net-worth client records and cross-referenced foundation and university financial statements to surface 30+ organizations with investable endowment assets.
Publications
Project Highlights
AI-powered personal career wiki
A multi-tenant AI career-wiki SaaS that turns raw career material into a structured knowledge base through Zod-constrained LLM extraction. It powers grounded RAG chat with inline citations, source-traceable resume generation, and voice mock interviews with LLM-as-judge scoring.
Engineered a sealed LLM gateway with atomic budget reservations, tiered model routing, cost auditing, Stripe billing, rate limiting, and server-side tenant isolation. Batched embeddings reduced ingest latency from 15 seconds to 600 ms.
Trading simulation platform & agent evaluation sandbox
Built a Node.js and TypeScript trading platform with PostgreSQL trade ledgers, position reconstruction, rule engines, and live market data. Open-sourced an npm MCP server with 16 tools for AI trading-agent evaluation, including fee/slippage-aware order execution, prop-firm risk rules, and replay backtests that prevent look-ahead bias.
Classical computer vision vs deep learning
Built an end-to-end MNIST pipeline comparing HOG with LinearSVC against a custom 1-channel ResNet18, with unified preprocessing, evaluation, metrics reporting, and an interactive canvas for real-time handwritten-digit predictions.
Achieved 97.87% test accuracy with HOG and 98.64% with ResNet18, a 0.77 percentage point improvement on the 60,000-image training and 10,000-image test dataset.
Skills
C++, Java, Python, Go, JavaScript, TypeScript, HTML/CSS
PyTorch, CUDA profiling, Node.js, React, Vue, Pandas, Spring Boot
Redis, MySQL, PostgreSQL, Docker, Bash, Git
Interests