The Mud King
中文

Builds, patterns & research

The LaboratoryThe Royal Laboratory

What has been built and is running, the patterns that keep sinking AI programmes, and the questions about agentic AI I am still working on.The things that survived contact with reality, the craters left by the ones that didn’t, and the questions I’m still poking with a stick.

  • In production
  • Running
  • Live on this site

Built and runningThings that survived contact with reality

Each entry is from a programme on the record. The figures are on the case-study pages.All real. All still running. Nobody had to be bribed.

In production2024 – present

Industrial machinery · Tianjin

Plant-edge computer vision for quality control

Can defects be caught on the line, by cameras at the plant edge, instead of by sampling afterwards?

What was built
Computer-vision quality control at the plant edge, part of the Tianjin Lighthouse-Factory programme alongside IIoT, MES, GenAI and digital twins.
Outcome
Near-perfect defect detection on the line. The set-up became the replicable global site template.

LessonPut the model where the defect happens, and design the template for the next site while you build the first.

  • Computer vision
  • Edge compute
  • MES
  • IIoT

In production2024 – present

Industrial machinery · global

Digit-AI use-case gates

How do you decide which AI ideas get funded, and move the good ones to production faster?

What was built
A framework connecting digital platforms, data architecture and AI models, with gates from concept to production.
Outcome
17.2% efficiency improvement across the China business. Now the basis of the group’s global use-case pipeline.

LessonMost of the value is in the gates, not the models. A use case that cannot name its cost or revenue never needs a model.

  • Use-case pipeline
  • MLOps
  • AI FinOps

Running2024 – present

Industrial machinery · China

APAC AI Innovation Lab

Can use cases proven in China become group standards instead of local one-offs?

What was built
An AI lab in China as the origin point for group-relevant use cases, including plant-edge computer vision and GenAI for supply-chain and consumer intelligence.
Outcome
China-proven patterns industrialised into the global site template and the global AI portfolio.

LessonA lab earns its keep through what leaves it. Count exits to production, not ideas generated.

  • GenAI
  • Computer vision
  • Supply chain

In production2024 – present

Industrial machinery · 60+ countries

One security baseline, 170+ deltas

Can one platform carry 170+ regulations if each jurisdiction is a delta on a shared baseline?

What was built
A Global Infosec Platform for IT security, OT/ICS protection and AI-driven security governance, with the jurisdiction-specific deltas automated.
Outcome
Compliance overhead down 40%, and zero major regulatory findings, including NERC CIP audits.

LessonCompliance scales as one product with local deltas, not as sixty local projects.

  • NIST CSF 2.0
  • IEC 62443
  • ISO 27001:2022

Running2024 – present

Industrial machinery · global

Security for the AI layer

What does security look like when the asset is a model, its training data and its prompts?

What was built
Security-by-design standards for the AI/ML platform: protection of training and inference data under GDPR- and PIPL-class regimes, red-teaming, adversarial testing and model monitoring.
Outcome
Part of the current mandate, aligned to industrial cybersecurity practice.

LessonPrompts and training data are data flows with a location and an owner. Govern them like any other.

  • Red-teaming
  • Model monitoring
  • NIST AI RMF

Live on this site2026

This site

Corporate Jargon Translator

Can a small rule set translate most corporate sentences, in English and Chinese, without calling a language model?

What was built
Two phrasebooks, English and Chinese, each with about fifty phrases and ten worked examples, plus an optional model fallback for anything the rules miss. It runs in the browser.
Outcome
Most sentences people type are handled by the rules alone, in both voices, with no network call.

LessonWrite the rules first. Call the model only for what the rules miss, and you pay for far fewer calls.

  • TypeScript
  • Phrasebooks
  • Optional LLM

Patterns that failHow AI programmes drown

Seen across programmes and industries. Recognising the pattern early is cheaper than recovering from it.Every one of these has a body count. Learn from other people’s craters.

  1. The permanent pilot

    What it looks like
    Dozens of pilots alive and none finished. The same demo is shown to every new executive.
    Why it happens
    No scale-or-stop date, and no agreed definition of “good enough to scale”.
    Do this instead
    Write down the number that must move and the decision date before kick-off. Make the call on the day.

  2. Automating the mess

    What it looks like
    A bot runs a twelve-step process faster, including the nine steps nobody needed.
    Why it happens
    The business case assumed the current process was worth keeping.
    Do this instead
    Remove steps first. What is left to automate is smaller, cheaper and easier to trust.

  3. The vendor’s data

    What it looks like
    A flawless demonstration, followed by a slow and expensive disappointment.
    Why it happens
    The proof ran on clean sample data, outside your security constraints.
    Do this instead
    Ask for a proof of value on your data, inside your perimeter, before the contract.

  4. Prediction before data

    What it looks like
    Predictive-maintenance alerts that operators learn to ignore.
    Why it happens
    Too few recorded failures to learn from, and gaps in sensor coverage.
    Do this instead
    Start with condition thresholds and better instrumentation. Predict once the history exists.

  5. Model before knowledge

    What it looks like
    Fluent answers that are confidently wrong about certifications, prices or specifications.
    Why it happens
    Retrieval over a knowledge base that nobody had cleaned or owned.
    Do this instead
    Fix the source content and give it an owner first. The model is the easy part.

  6. One fast step in a slow process

    What it looks like
    One step is automated and the cycle time does not move.
    Why it happens
    The delay sits somewhere else, often in an approval chain.
    Do this instead
    Map the whole process end to end and fix the bottleneck, wherever it is.

Research agenda: agentic AIQuestions I can’t stop asking

Open questions about landing agents in industrial, regulated and cross-border settings. Partners welcome.Nobody has these answers yet, including the vendors. Especially the vendors.

  1. How reliable must an agent be before it may act without a person?

    Why it matters
    The answer differs by process and by risk, and most organisations have no threshold written down.
    How I’m approaching it
    Risk-tiered pass marks built from real evaluation sets: what error rate is acceptable for a reversible action, and what for an irreversible one.

  2. What must an agent’s audit trail show?

    Why it matters
    The EU AI Act, ISO/IEC 42001 and China’s rules on generative AI all expect traceability, in different words.
    How I’m approaching it
    One trail format (inputs, tools called, decisions, human interventions) mapped against each regime’s requirements.

  3. Where should agents run when the data cannot leave the country?

    Why it matters
    Data residency under PIPL and similar regimes limits which models and clouds an agent may use.
    How I’m approaching it
    Comparing deployment patterns, from domestic models and clouds in China to sovereign and hyperscaler options elsewhere, and how one agent design can span both.

  4. Who holds authority: the agent, or the system of record?

    Why it matters
    When agents act inside ERP, MES and CRM, it becomes unclear who approved what.
    How I’m approaching it
    Treating agents as users of the systems of record, with their own identities, permissions and approval limits.

  5. What does an agent really cost per task?

    Why it matters
    Inference, tools, oversight and rework land in different budgets, so the real cost is rarely visible.
    How I’m approaching it
    Extending AI FinOps from cost per call to cost per completed task, including the human time it saves and the human time it uses.

  6. Which industrial processes are ready for agents?

    Why it matters
    Agents need digital inputs, clear rules and reversible actions. Many processes have only one or two of the three.
    How I’m approaching it
    A readiness screen built on those three tests, applied to quotation, maintenance, quality and supplier processes.

Working on one of these? Request an audience