Skip to content
TypeScriptNode.jsAI AgentsDeveloper ToolingAutomation

Building tools that make AI actually work for developers

Ming — Software developer building autonomous systems, developer tools, and AI-powered workflows.

See my work

Now

Building a knowledge management system to power business in the agentic era. Leading AI adoption at GBST — building protocols, tools, and workflows that make autonomous development reliable.

45 → 5 min

CVE triage time reduced per vulnerability through AI-powered scanning

8 hrs/week

Recovered through AI workflow automation

15+ tools

Benchmarked against open-source AI coding tools (800k+ combined stars)

7 repos

Public AI tools on GitHub — dev-kit, auto-research, llm-router, and more

Skills

AI / ML

Claude CodeGPT-4oDeepSeekVercel AI SDKPaperQA2Semantic Scholar

Languages

TypeScriptPythonShellRust

Frameworks

Next.jsReactNode.jsSpring FrameworkFastAPI

Infrastructure

AWSKubernetesDockerCDPsystemd

Data

PostgreSQLSQLiteDrizzle ORMRedis

Testing

VitestPlaywrightaxe-core

Agentic AI Era

Updated Jun 2026

dev-kit

AI-native development toolkit — methodology as code

Situation

AI coding agents can write code, but they implement the wrong thing, game their own tests, skip quality gates, and don't scale beyond single-agent workflows.

Decision

Built a methodology-as-code system: spec-driven development with EARS notation, information barriers (coders never see specs), hidden regression tests (40% invisible to coders), role-separated multi-agent orchestration, and file-based pipeline FSM.

Outcome

A portable toolkit that bootstraps any project into a rigorous AI-assisted development environment in under a minute. Clone, run scaffold.sh, start building.

13 agent roles · 14 skills · 11 workflows · benchmarked against 15+ open-source AI tools

TypeScriptShellClaude CodeVitestPlaywright
View on GitHub

auto-research

Autonomous overnight research engine — idea seeds to papers

Situation

Research takes hours of manual searching, reading, and synthesizing. Most people never do it systematically, and valuable ideas stay unexplored.

Decision

Built an autonomous engine that takes idea seeds and produces research papers overnight — searching Semantic Scholar, arXiv, and university libraries with quality gates and integrity checks.

Outcome

Wake up to enriched feasibility reports with academic-quality evidence, structured bibliography, gap analysis, and Typst-formatted papers ready for compilation.

Nightly cron (02:00–04:45) · 3 academic sources · DEER quality scoring · Integrity checks

TypeScriptoclifPaperQA2Semantic ScholararXivTypst
View on GitHub

llm-router

Free-tier LLM gateway — 11 providers, one OpenAI-compatible API

Situation

Free LLM tiers are fragmented across 10+ providers with different APIs, quotas, and availability. Switching between them manually is tedious and unreliable when providers hit rate limits or go down.

Decision

Built a gateway that aggregates 11 free-tier LLM providers behind a single OpenAI-compatible API with availability-aware smart routing, quota management, and circuit breakers. Health probes every 2h, model discovery every 6h. Task classification routes code/reasoning/summarize/general requests to the best available model.

Outcome

One API endpoint that automatically routes to the best available free-tier provider. Circuit breakers with DB persistence survive restarts. Batch processing for overnight jobs. Drop-in replacement for any OpenAI SDK.

11 providers · Smart routing · Circuit breakers · OpenAI drop-in

PythonFastAPISQLitesystemd
View on GitHub

agent-workstation

Orchestration platform for AI coding agents — multi-agent session management

Situation

AI coding agents like Claude Code work in isolation — no session persistence, no agent-to-agent coordination, no way to manage multiple agents working on different parts of the same codebase simultaneously.

Decision

Built an orchestration platform that manages agent sessions with workspace isolation, shared context through file-based communication, and a session lifecycle that supports spawning, monitoring, and merging agent work. Designed for multi-agent workflows where different roles (coder, tester, reviewer) operate on the same project.

Outcome

A unified platform for running multiple AI coding agents in coordinated sessions. Agents can work in parallel on isolated worktrees, share context through structured files, and merge results back. Session state persists across restarts.

Multi-agent sessions · Worktree isolation · Shared context · Session persistence

TypeScriptNode.jsShell
View on GitHub

browser-cli

3-layer browser automation — CDP, accessibility trees, AppleScript

Situation

Browser automation is fragile — CDP connections break, sites block headless browsers, and OAuth flows require visible browser sessions.

Decision

Built a 3-layer architecture: thin CLI router that delegates to an accessibility-tree-first engine, backed by a raw CDP SDK with 652 typed wrappers. AppleScript fallback via SSH for OS-level operations on macOS.

Outcome

One unified browser command that works across WSL2, macOS, and bot-blocked sites. Persistent sessions, structured output, and automatic environment detection.

40+ commands · 3 environment modes · 56 CDP domains · persistent sessions

TypeScriptBunCDPAppleScriptRust

LiveCase Tutor Bot

AI research platform — studying AI-assisted decision-making

Situation

Business school students use strategic management simulations, but there's no way to study how different AI tutoring personas affect their decision quality, speed, and confidence.

Decision

Built a Chrome extension + Next.js 14 research platform with a 3×3 factorial A/B test design: 3 AI personas (Strategic Architect, Stakeholder Provocateur, Sustainability Alchemist) × 3 prompt variants. Implemented behavioral data collection tracking decision changes, hesitation patterns, and AI influence.

Outcome

Production research platform deployed at hs.mingli.world with 5 analytics endpoints: engagement, AI influence measurement, variant comparison, group comparison, and CSV export for statistical analysis.

3×3 factorial design · 5 research endpoints · Qualtrics integration · production deployment

Next.js 14Plasmo Chrome ExtensionPostgreSQLDrizzle ORMVercel AI SDKGPT-4o

More Projects

nexus

Knowledge engine SDK — ingest, process, serve with LLM pipelines

Pipeline SDK · Zod validation · Vector search · MCP server

TypeScriptZodSQLitep-queue
GitHub

ai-feeds

AI industry intelligence — 7 sources scored against your learning plan

7 sources · 332 tests · LLM scoring · Obsidian digests

TypeScriptSQLitePlaywrightClaude
GitHub

generic-tutor

AI tutor engine with SM-2 spaced repetition

SM-2 algorithm · Knowledge DAG · Agent-loadable

ShelljqMarkdown
GitHub

Education

Master of Science, Financial Engineering

The Chinese University of Hong Kong

2016

Master of Science, Computer Vision

University of Chinese Academy of Sciences

2015

Bachelor of Science, Automation

Tsinghua University

2011

Let's build something together

Open to Software Developer and AI Engineer roles. Let's talk.