Work & Research

I build agents for complex workflows and co-author research on structured generation, precise editing, and tool use. Here are the problems, approaches, and public sources behind that work.

Agent engineering

Nova Agent

Production agents · Insurance

Problem. Insurance operators work across large document sets and internal services to prepare contracts and communications.

My work. I work on agent system design and production engineering around tools, skills, memory, document processing, permissions, and human review. Work on quoting and placement services informs how I integrate the agent with the surrounding workflow.

Approach. The workflow connects policy and submission documents to generated artifacts that operators can inspect. A single task can involve more than 50 PDFs.

Read the project background

TikTok coding agent

Developer tooling · Code migration

Problem. Migrating Objective-C code to Swift requires generated code that is useful within an existing engineering workflow.

My work. I built the internal agent from scratch, including code-generation loops and custom evaluation harnesses, and iterated on prompts and generated code.

Approach. Evaluation was part of the development loop: generated code was checked and refined for practical migration work. My mobile product engineering experience helped ground the tooling in developer needs.

Read the project background

Selected research

FineEdit

Findings of EMNLP 2025

Problem. Instruction-driven edits must preserve surrounding content and respect the structure of code, LaTeX, and database languages.

Research. Co-authored FineEdit, connecting a specialized editing model with the InstrEditBench benchmark for targeted modifications.

Approach. The research evaluates precise edits across structured domains and extends evaluation to multi-turn editing scenarios.

Read the paper

TreeDiff

SURGeLLM at ACL 2026

Problem. Random token masking in diffusion language models does not capture the syntactic boundaries and hierarchical dependencies of code.

Research. Co-authored TreeDiff, a syntax-aware diffusion framework that uses abstract syntax tree structure in the corruption process.

Approach. The method masks tokens belonging to selected AST nodes, aligning reconstruction with program structure. Published in the SURGeLLM workshop proceedings at ACL 2026.

Read the paper

Infant Agent

Agent research · arXiv preprint, 2024

Problem. Multi-step engineering and reasoning tasks need coordination beyond a single model response.

Research. Co-authored research on a tool-integrated, logic-driven agent architecture with task-aware functions and hierarchical management.

Approach. The work studies memory retrieval, tool integration, and API costs, with evaluations on software engineering and mathematical reasoning tasks.

Read the paper

Earlier research

RED

Real-time scheduling for robotic inference under changing environmental dynamics. Published at RTSS 2023.

Web accessibility research

Constraint solving and optimization for finding and repairing accessibility issues in web applications, with Yi Liu and Prof. Yepang Liu's lab.

Experience

Newfront / WTW

October 2025 - Present

Insurance AI systems, Nova Agent, and quoting and placement services.

Tira AI

Summer 2025

Co-founded an AI-native execution startup with Sam Liu; worked on product architecture, implementation, and early go-to-market development.

TikTok

Before October 2025

Consumer mobile product systems and an internal LLM coding agent for Objective-C to Swift migration.

Ansys

May–August 2022

Digital Twin internship: frontend, APIs, and browser-compatible deployment workflows for Twin Deployer.

University of Pittsburgh

Research beginning in 2021

Web, mobile, data collection, visualization, and low-code rehabilitation research platforms in Prof. Dan Ding's lab.

Ant Group / Alibaba

July 2020–June 2021

Alipay financial product systems, low-code infrastructure, live-streaming components, and backend services.

SUSTech

2016–2020

Computer science degree, GAN-based face generation with Prof. Ran Cheng, and web accessibility research.

Education

  • University of Pittsburgh — MS, Information Science, 2021–2023
  • Southern University of Science and Technology — BEng, Computer Science and Technology, 2016–2020

Complete résumé · Selected projects · Contact Tempest