Building reliable agents for real-world work.

I build systems that connect model capabilities to tools, documents, and the constraints of a real workflow. My interests include structured code generation, precise editing, evaluation, and human review. I care about what happens after a model produces an answer: whether the result is useful, testable, and reliable enough to use.

Currently building

Nova Agent · Newfront / WTW

An internal AI coworker for insurance operators. I work on document-heavy workflows, generated contracts, company-style email drafts, tool use, and human-in-the-loop review.

A task can involve more than 50 PDFs alongside policy and submission documents. The engineering challenge is to coordinate those files and internal services while respecting permissions, domain constraints, and operator review. My work on quoting and placement services also informs how I design the agent around the insurance workflow.

Selected work

TikTok coding agent

Developer tooling · Code migration

Built an internal LLM agent for Objective-C to Swift migration, with generation loops and custom evaluation harnesses.

FineEdit

Findings of EMNLP 2025

Research on precise, instruction-driven editing in structured domains including code, LaTeX, and database languages.

TreeDiff

SURGeLLM at ACL 2026

AST-guided code generation with diffusion language models, treating code as a structured object.

Explore all work and research

A path through research and engineering

I began experimenting with deep learning applications in 2014. At SUSTech, work on GAN-based face generation and web accessibility introduced me to research problems that combined models, formal structure, and practical engineering.

At Ant Group, I built software for financial workflows. At Pittsburgh, I worked alongside rehabilitation researchers on systems for studies and training programs. My TikTok work connected mobile engineering with LLM-based developer tools. Those experiences now inform how I build agents for insurance operations.

I have also built products from the ground up, including Tira AI and AgentShelf. Building for users keeps my research questions grounded: what needs to be evaluated, where does the workflow break, and what should remain under human control?

Read my complete research and engineering journey · View projects

Questions I keep returning to

• Reliable tool-using agents
• Structured generation and precise editing
• Evaluation under real constraints
• Human-in-the-loop workflows
• Real-time AI systems

For me, useful intelligence includes the system around the model: tools, memory, evaluation, permissions, and runtime behavior. I am interested in how these pieces work together when documents are messy, tasks span many steps, and people need to inspect the result.

Writing

• From Deep Learning to Production AI Agents
• Civilizational Intelligence
• My Frontend Journey

Browse all writing · Technical background

Let's connect

Interested in research, agent systems, or engineering collaboration? Email me or find me on LinkedIn and GitHub. For career opportunities and current availability, please contact me directly.