Life Science × Local AI Systems

I build verifiable AI tools for real research and development.

I work across genetics, developmental biology and bioinformatics, and turn recurring AI-workflow problems into local-first, inspectable tools.

View Projects Build Log

01 / Selected work

Turning real problems into verifiable systems.

From bounded coding agents to a local model gateway and searchable working memory, each project starts with a recurring problem and ends with something inspectable.

Active development

AI AgentCode ReviewWorkflow

Making AI agent output auditable code changes

Bridging ChatGPT solution discussion with local coding agent execution — returning status, changed files, verification, audit, and task lineage evidence within a bounded workspace.

View Details

RelayForge

Active development

Zero-dependency, local-first AI coding gateway compatible with OpenAI/Anthropic APIs, supporting combo routing, fallback, privacy redaction, and lightweight usage analytics.

CodexJournal-Lite

Active development

Archive local records from Codex, Claude Code, Gemini CLI, OpenCode, JetBrains/IDEA — producing structured Markdown/JSON output, full-text search, and a localhost-only dashboard.

XJTU Codex Theme

Original

Windows local theme project for Microsoft Store Codex Desktop — includes XJTU Academic Theme, Codex Skin Maker, and Unbranded Template. Source-available under a non-commercial license.

View All Projects

02 / Field notes

Build Log

Field notes on local agents, verification, and the failures that changed how I build.

Read all notes

2026-02

The Real Cost of Vibe Coding & Verification

AI-assisted programming is not zero-cost programming. The verification stage determines the reliability of output.

2026-01

ChatGPT × Local Agent Collaboration

Two types of AI tools each have strengths and weaknesses. Finding the collaboration pattern between them is key.

03 / Method

Method matters more than speed.

Define the problem and boundaries first, let agents implement, then verify the result and preserve evidence that can be reviewed later.

Discuss

Clarify requirements, discuss approaches, define boundaries in ChatGPT

Plan

Break down tasks, define milestones, limit workspace scope

Execute

Have local agents (Codex, Claude Code) implement according to plan

Verify

Review diffs, run tests, confirm output meets expectations

Audit

Retain task trail evidence for traceability and review

04 / Contact

Let’s discuss a problem worth solving.

If you work around life science, controllable agents, or local-first tools, I’d be glad to compare notes.

Send Email GitHub