Building tools that make AI actually work for developers
Ming — Software developer building autonomous systems, developer tools, and AI-powered workflows.
See my workThe thesis
Everyone is about to own a software producer business. The infrastructure they will need is not another agent — it is a company.
Building software has stopped being the bottleneck. Agent harnesses like OpenClaw and Hermes solved how agents run and how they learn, and a person with a good idea can now produce working software faster than they can decide what to charge for it.
What nobody has solved is everything that turns that output into a business. Who is allowed to do what. What happened, and who authorised it. What each run cost and what it earned. Which work is queued, which is blocked, and what broke overnight. Run one agent and you can hold that in your head. Run a crew, and you need a company.
Foundry is that company operating system: identity and RBAC, an append-only audit trail, durable task dispatch, cost accounting and a BI event stream — with the general-purpose services on top (crawling, knowledge, matching, an agent runtime) that applications consume through typed contracts.
I am building it the only honest way to build it: by running my own one-person company on it. Every application below is a real product with real users' problems, and each one onboards onto the platform. If the platform cannot carry them, the thesis is wrong and I will know before anyone else does.
OpenClaw and Hermes gave agents a runtime. Foundry gives them a business to run.
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.
133k LoC
CompanyOS platform — 4 apps, 4 services, 11 packages, 28 architecture decisions
45 → 5 min
CVE triage per vulnerability, via LLM-powered scanning shipped at GBST
80%
Reduction in developer hours spent on vulnerability work
12 products
Applications and toolchain onboarding onto the platform
Skills
AI / ML
Languages
Frameworks
Infrastructure
Data
Testing
⚡ The stack
Updated Jun 2026The platform
The company operating system everything else runs on.
Foundry — CompanyOS
The company operating system for a one-person AI-native business
Situation
Running a crew of agents surfaced the real gap: the agents worked, the company around them did not. Nothing recorded who authorised what, what each run cost, or which work was blocked.
Decision
Built a platform, not a script. Two planes on shared services — a product plane customers touch, a company plane the founder and the agent crew touch — so every capability is built once and every crew action is dispatched, budgeted and audited.
Outcome
133k lines of TypeScript across 4 apps, 4 services and 11 packages, governed by 28 architecture decision records and a 16-check verification gate. When that gate turned out to be grading a stale build, the fix and the failure were both written down.
133k LoC · 28 ADRs · 117 tickets
What runs on it
Products onboarding onto the platform, each solving a real problem first.
job-hunter
Job-hunting suite — sourcing, ranking, referral intelligence
Situation
Job boards return volume, not fit, and the highest-signal roles are hidden behind recruitment agencies.
Decision
13 sources feeding a staged funnel: eligibility filter, BM25 retrieval, LLM triage, then a judge that must quote the posting verbatim and is blocked from inventing tenure it was never given.
Outcome
Surfaces Sydney AI roles with referral routes mapped from a real LinkedIn session, and refuses to put a number on a resume that is not in the candidate's own documents.
13 sources · 12k listings · 249 tests
generic-tutor
Career-driven learning engine
Situation
Skill gaps found during a job hunt are forgotten by the time there is time to study.
Decision
SM-2 spaced repetition over interview drills, wired directly to the skill gaps job-hunter and ai-feeds surface.
Outcome
Learning targets come from the market and the candidate's real gaps rather than a generic syllabus.
SM-2 · interview drills
trader
Trading signal agent
Situation
Market opportunities need watching continuously; a person cannot.
Decision
An agent that monitors instruments and notifies on conditions worth a human decision.
Outcome
Notification on opportunity rather than continuous attention.
always-on
digest-assistant
Personal automation hub
Situation
Personal signal is scattered across mail, feeds and repositories.
Decision
Six pipelines and a knowledge graph behind a single Telegram bot.
Outcome
One daily surface for everything that would otherwise need six check-ins.
6 pipelines · 1 bot
ai-feeds
AI industry intelligence
Situation
Keeping current with papers, repos and releases is a full-time job.
Decision
Aggregates signals and converts them into actionable learning artifacts and upgrade ideas for existing repos.
Outcome
Feeds the auto-research pipeline, which turns ideas into researched plans overnight.
papers · repos · releases
case-study
Case-based learning with an AI mentor
Situation
Decisions under incomplete information cannot be taught from a textbook.
Decision
Learners run a business scenario as the decision-maker while an AI mentor probes the reasoning.
Outcome
Also the platform behind university research into AI-assisted decision-making.
research platform
What builds it
The agent toolchain the platform and its applications are built with.
dev-kit
AI-native development toolkit — methodology as code
Situation
Agent quality is bounded by the process it is given, and the process lived in one person's head.
Decision
23 skills, 6 hooks, spec-driven development and multi-agent orchestration, written entirely in markdown and bash so it is portable across harnesses.
Outcome
The same methodology drives every project in the stack.
23 skills · 6 hooks
agent-workstation
Orchestration platform for AI coding agents
Situation
Multiple agents on multiple terminals need to cooperate without stepping on each other.
Decision
Session management and orchestration across concurrent agents.
Outcome
Agents on separate terminals coordinate through a shared channel.
multi-agent
kiro-sessiond
Session daemon for agent orchestration
Situation
Concurrent agents editing the same repository corrupt each other's work.
Decision
Agent registry, scoped message queue, file claims and an HTTP API.
Outcome
File claims make concurrent agent work safe.
registry · claims · queue
nexus
Knowledge engine SDK
Situation
Every application needed the same ingest-process-serve pipeline.
Decision
A library, deliberately not a platform — embedded by foundry and job-hunter, extended upstream.
Outcome
One knowledge engine behind several products.
embedded by 2 products
llm-router
Free-tier LLM gateway — 11 providers, one API
Situation
Provider limits, outages and quotas break anything built on a single model.
Decision
Availability-aware routing across 11 providers behind one OpenAI-compatible API, with quota management, circuit breakers and batching.
Outcome
Applications target one endpoint; the router handles provider reality.
11 providers
knowledge-hub
Knowledge graph engine with conflict detection
Situation
Issue trackers, repositories and specs disagree, and nothing notices.
Decision
Adapter-based graph that connects all three into a queryable context layer and flags contradictions.
Outcome
Conflicts surface before they reach an agent as context.
conflict detection
More Projects
nexus
Knowledge engine SDK — ingest, process, serve with LLM pipelines
Pipeline SDK · Zod validation · Vector search · MCP server
ai-feeds
AI industry intelligence — 7 sources scored against your learning plan
7 sources · 332 tests · LLM scoring · Obsidian digests
generic-tutor
AI tutor engine with SM-2 spaced repetition
SM-2 algorithm · Knowledge DAG · Agent-loadable
Education
Master of Science, Financial Engineering
The Chinese University of Hong Kong
Master of Science, Computer Vision
University of Chinese Academy of Sciences
Bachelor of Science, Automation
Tsinghua University