AI, Engineering & Systems

Problems solved

LLM outputs were entering a permanent record unverified.

I built an AI-operable research workspace — an MCP server and CLI that lets an AI agent search, ingest, and sync papers across Zotero, Obsidian, and NotebookLM — with references that cannot be verified quarantined under an explicit failure taxonomy rather than silently accepted. Shipped and maintained publicly as research-hub-pipeline on PyPI.

Generative agents entered consequential simulation with no governance standard.

I co-authored the Water Resources Research (2026) study substituting LLM agents into a human-water simulation, and now build the evaluation methods that test whether such agents reproduce empirically measured human decisions — with ground truth I collected myself (a 937-household survey).

Catastrophe loss estimates are only as good as their validation.

I built a flood-loss pipeline that publishes its validation honestly — a 52,141-household coupled agent-based and catastrophe model with NFIP financial mechanics, validated against observed insurance claims — and worked on a team-built Hazus-based seismic bridge-loss pipeline validated at three levels against the 1994 Northridge earthquake, with failures reported, not hidden.

Systems

research-hub

Maintained open-source tool

AI-operable research workspace: a CLI, MCP server, REST API, and dashboard that drive repeatable literature workflows across Zotero, Obsidian, and NotebookLM.

PyPI research-hub-pipeline v1.1.1; listed in the Awesome MCP Servers catalog; tested on Windows, macOS, and Linux; an automated test suite that checks retrieval quality.

Case study Repo

FLOODABM

Research prototype · Archived companion code

52,141-household coupled ABM × catastrophe flood model, Passaic River Basin, NFIP mechanics, calibrated on a 937-household survey I designed and fielded.

Public repo; Zenodo archive with CITATION.cff and published seed lists; known-limitations register.

Case study Repo

Cat_framework

Research prototype · Team capstone

Team-built Hazus 6.1 seismic bridge-loss pipeline (hazard → exposure → fragility → EP/AAL).

Public repo; three-level Northridge validation with honestly reported failures.

Case study Repo

codex-delegate

Maintained open-source tool

Cross-platform agent-delegation wrapper whose completion claims are verified from git state.

Public repo; Ubuntu/Windows CI; regression test pinning a real upstream stdin hang.

Case study Repo

awesome-agentic-ai-zh

Maintained open-source tool

Trilingual agentic-AI curriculum (community/education line).

4.7k+ GitHub stars, 622 forks (July 2026); external contributors; automated checks keep all three languages in sync.

Case study Repo

WAGF — a governance framework for LLM agents in human–water simulation, keeping coupled behavioral outcomes stable and physically sensible; research software in preparation for release.

Evaluation & validation record

research-hubIn-repo retrieval-recall evaluation suite against golden fixtures; observed failures recorded as explicit audit baselines.
research-hubA fail-closed anti-fabrication module with a transient-vs-permanent failure taxonomy.
codex-delegateA regression test that reproduces a real upstream failure mode and fails without the fix.
Cat_frameworkMulti-level hazard-model validation with honestly reported misses.
FLOODABMSimulation validation against observed insurance-claims data.

Systems designed so that failure is visible — no silent passes.

Open-source contributions

18 merged pull requests in third-party open-source projects.

BuilderIO/agent-native#1332Documentation fix in the development guide.
BuilderIO/agent-native#1333Documentation-reference fixes across contributor docs.
BuilderIO/agent-native#1334Documentation fix removing broken links from a skill guide.
BuilderIO/agent-native#1362Root-caused engine-selection bug fix with a regression test that fails on main.
Nanako0129/coralline#10Pure-bash UTF-8 / East-Asian-width display fix.
langchain-ai/openwiki#367Security fix restricting ~/.openwiki permissions on Windows, where chmod is a no-op.
ZhuLinsen/daily_stock_analysis#1773Taiwan market detection and routing.
ZhuLinsen/daily_stock_analysis#1801Made Taiwan a first-class market across the service layer, API schema, and typed frontend.
ZhuLinsen/daily_stock_analysis#1829Taiwan institutional-flow (foreign, trust, and dealer) fetcher over official TWSE/TPEx open data, fail-open.
NVIDIA/skills#76Documentation fix repairing the Skill Catalog link in the contributing guide.
ZhuLinsen/daily_stock_analysis#1841Coalesced concurrent institutional-flow fetches so a cold cache cannot trigger a stampede of duplicate requests.
ZhuLinsen/daily_stock_analysis#1855Network-marked drift tests for the institutional-flows fetcher, so an upstream format change fails loudly instead of silently.
ZhuLinsen/daily_stock_analysis#1863Surfaced Taiwan institutional flows in the report's institution block.
ZhuLinsen/daily_stock_analysis#1864Hardened the institutional fetcher with a circuit breaker and an exchange-date guard.
ZhuLinsen/daily_stock_analysis#1866Surfaced Taiwan institutional flows in the report and the LLM prompt, and corrected TWD currency handling.
ZhuLinsen/daily_stock_analysis#1867Fixed a dividend parse that made the trailing-twelve-month yield silently wrong.
ZhuLinsen/daily_stock_analysis#1869Detect the Taiwan closing-auction window (13:25-13:30) as a distinct market phase.
punkpeye/awesome-mcp-servers#6921Added research-hub to the Research category of the Awesome MCP Servers catalog.

Capabilities

Python · Testing and CI · Evaluation design · MCP and agent tooling · Git · Cross-platform (Windows/POSIX) engineering · Catastrophe-risk stack: FEMA Hazus, fragility functions, exceedance-probability curves, NFIP mechanics