Ph.D. Candidate, Civil & Environmental Engineering, Lehigh University
AI-agent infrastructure, and the science of when to trust it
AI evaluation / research engineer with catastrophe-risk domain depth.
I build AI-agent infrastructure — MCP tooling, multi-agent orchestration, and agent harnesses — and I apply large language models as the behavioral engine of agent-based models of complex human–water systems, grounded in survey data I collected myself, with the governance that keeps agent behavior trustworthy.
Open to Summer 2027 internships in AI evaluation, research engineering, and risk analytics.
Shipped AI infrastructure
research-hub — an MCP server and CLI (PyPI: research-hub-pipeline v1.1.1) that makes Zotero, Obsidian, and NotebookLM AI-operable. Listed in Awesome MCP Servers; tested on Windows, macOS, and Linux.
Maintained open-source tool
AI-agent skills & governance
Open-source agent skills — ai-research-skills (a 15-skill research catalog), agent-collab-skills, and codex-delegate — plus WAGF, a governance framework that constrains LLM agents so agent-driven simulations stay trustworthy.
Maintained open-source tools · WAGF in preparation
Community & education
awesome-agentic-ai-zh — a trilingual roadmap for building agentic AI, from LLM basics to production multi-agent systems, with automated checks that keep all three languages in sync.
4.7k+ GitHub stars, 622 forks (July 2026)
Maintained open-source tool
Literature workflows are repetitive and easy to get subtly wrong across tools. research-hub turns Zotero, Obsidian, and NotebookLM into one AI-operable workspace — searching, ingesting, and syncing papers through a single CLI, MCP server, and REST API.
Maintained open-source tool
Delegating code work to a second AI agent is only economical if the delegate cannot fabricate success. codex-delegate verifies the delegate’s completion from git state rather than from its own report.
Maintained open-source tool
Chinese-speaking learners lacked a staged path into agentic AI, and hand-mirrored translations drift. This trilingual curriculum keeps its locales verifiably in sync with CI.
4.7k+ GitHub stars (July 2026)
Research prototype
Flood-adaptation models usually assert household behavior instead of measuring it. FLOODABM grounds 52,141 simulated households in a 937-household survey and validates losses against observed insurance claims.
Research prototype
A loss number without an inspectable validation chain cannot be trusted. This Hazus-based seismic bridge-loss pipeline reports where the recipe fails, not only where it works.
Current focus: evaluation and governance methods for LLM agents — testing whether they reproduce real human decision pathways (grounded in a 937-household survey I designed and fielded) and constraining agent-driven simulations so they stay trustworthy. Research software in preparation for release.
Open to Summer 2027 internships in AI evaluation, research engineering, and risk analytics.
wec324@lehigh.edu