Work

Things I've made.

Six things I have built, each a different kind of thing, then the full record. Rest on a piece and its sheet is inked.

01

Built

Countersign

A way to let AI agents do real work without losing control of it. Before an agent can take an action, a small rule-based gate outside the agent checks it: how risky is this, who is the named person responsible, and is this action on the list at all. Routine work passes straight through. Anything serious pauses until the right person signs it off. Anything unknown is refused. Every decision is written to a log that cannot be quietly edited. The rules for your business, your country and your roles plug in as packs, so the core never has to change.

Why it mattersAn agent can be talked into things. A gate cannot. If you want AI to act on your behalf, this is how you stay accountable for what it does.

02

Running

The second brain

A shared memory for a company, built so that the right people see the right things. Every note, message or document is labelled as it comes in: who it belongs to, how sensitive it is, where it may be used. Anything it cannot label is held aside rather than dropped. When someone asks a question, the answer is filtered by who is asking before the AI ever sees it. Overnight, the system reads what was added and suggests new facts; a person accepts or rejects each suggestion from a card that shows its reasoning. Deleting someone's data is a proper action with a record, not a request that gets lost.

Why it mattersYour team's knowledge becomes searchable without the newest hire seeing the payroll file. The AI helps, and a person stays in charge of what counts as true.

03

Whitepaper, in draft

Governing Agentic AI by Construction

The thinking behind Countersign, written down properly. It explains why governing AI that acts is a different problem from governing AI that advises, and sets out an arrangement where the agent does the work and a named person stays accountable, enforced by code rather than by policy. Every promise the framework makes is stated, tied to the control that enforces it, and backed by a test that fails if that control is removed. It comes with a one-page brief, a guide to Singapore's rules, and a short self-check to run before you deploy an agent.

Why it mattersMost governance documents are written to be filed. This one is written to be tested, and it is honest about which parts are finished and which are still to come.

04

Runs every engagement

The engine

A complete method for building a brand and the operations behind it, turned into software that actually runs it. The work is broken into hundreds of small steps, each with a clear owner, input and output. Between the steps sit gates: points where the work is checked and signed off before anything moves on. Nothing skips a gate. Every decision made along the way is recorded, and what is learned on one engagement is kept for the next.

Why it mattersYou get the same rigour whether it is the first engagement or the fiftieth, and you can see exactly why every decision was made.

05

Shipped

RTSN Studios, the app

A creative studio that runs entirely on your own computer. Generate images, pull a picture apart into editable layers, read the depth of a scene, and talk things through with a local AI model. Each job has its own room. The studio can also be handed to an AI assistant as a set of tools, so it can work the rooms for you. Nothing has to be uploaded anywhere.

Why it mattersNo fee per picture, no waiting on someone else's servers, and your work never leaves the building.

06

Verified on the device

Firmware on the bench

Small physical things, programmed from scratch. A concert wearable worn upside-down on a lanyard, with a pet, a countdown, a setlist that buzzes in time with the crowd. An LED frog that sleeps until you walk up to it and blushes when a hand comes near. A creature on a palm-sized screen that listens and talks back. Each one was checked on the real device, not in a simulator, because hardware has a way of disagreeing with the screen.

Why it mattersThe same habit that makes hardware work makes software work: test the real thing, in the real conditions, and believe what it tells you.

The full record

Everything else, grouped by what kind of thing it is: governance and memory, the studio, production systems, agents that run on their own, firmware, strategy, operations. The label on each line is its state as recorded at the time.

Governance and memory

  • Countersign: the governance core: a rule-based gate that checks every action, a named person behind each one, a pause for the serious ones, a log that cannot be quietly edited, and plug-in packs for different businesses and countriesBuilt
  • The second brain: a shared company memory where everything is labelled on the way in, answers are filtered by who is asking, and a person approves what the system learns overnightRunning
  • The whitepaper and its companions: Governing Agentic AI by Construction, with a one-page brief, a guide to Singapore's rules, a self-check before you deploy an agent, and a short presentationIn draft
  • The ontology: the engine's own memory: every step, gate and decision from every engagement, and the patterns drawn from themMaintained
  • Governed on-box AI briefings: an AI that reads a company's records and writes briefings for its owner; it recommends, never acts, and runs on a machine inside the buildingLive
  • Access control and privacy compliance: one list of who people are, one place to set what they can see, and a mapping of Singapore's AI guidelines onto how the system is checkedLive

Brand and studio

  • The engine: the full brand and operations method, run by software, with gates that cannot be skippedRuns every engagement
  • RTSN Studios, the app: a local creative studio with a room for each job, which an AI assistant can also operate as a set of toolsShipped
  • The v5 brand and this site: the palette, the type, the line Keep building, two pixel characters, a street that loops, and the drafting table you are readingLive
  • Four worked examples: fictional brands run through the engine from start to finish and published in full: a law chambers, a property agency, a restaurant group, a wealth managerLive, fictional
  • A Singapore tea brand: strategy, voice and identity taken through to a brand book, with a seal drawn from the name as its markDelivered
  • Client Portal: lets a client comment directly on the design itself, with access limited to themCode complete
  • Image and video pipelines: pictures and clips generated in-house, checked against the brand before they ship, signed so their origin can be proved, and packaged for each channel with captionsBuilt
  • Brand-system handoff: a standard package for leaving a client with a brand they can run themselves: the book, the design tokens, the assets and the rulesUsed twice

Production systems

  • Sales CRM and conversations: deals, enquiries and contacts in one place, with WhatsApp and email in a single inbox where a bot and the sales team work the same thread; replies to email are always drafted for a person to checkLive
  • Reporting estate: reads a mailbox and a drop folder every day, builds the tables, checks them, and publishes dashboards for different levels of accessLive
  • Scheduled sales reporting: the weekly and monthly workbooks produced and emailed on a schedule, proven to match the hand-built originals month for monthLive
  • One Axis: the reporting estate rebuilt around a single timeline of what has happened and what is expectedFirst phase live
  • Digest views, filters, rate metrics, a quotation tool, an enquiries connector, an analytics frontendLive or merged

Autonomous agents

  • Enquiries daily stand-up: reads a shared sales mailbox each morning and writes the stand-up: who is waiting on us, what moved, which promises are unkept, what new business arrivedRuns weekday mornings
  • Morning email routine: goes through yesterday's mail and calendar, drafts replies that are never sent without a person, and keeps track of what is still owedRuns daily
  • Pricing review and paid-search review: two advisory tools colleagues run themselves; they suggest and sanity-check, they never change anythingIn use
  • The daily brief: reads the morning's sources and hands back the handful of things worth knowingRuns daily
  • InvestOS: investment research with discipline built in: it looks for edges, argues both sides, keeps a journal that cannot be rewritten, and never places the orderRuns daily and weekly
  • Readable-HTML skill, design-system plugin: the brand as code: the design tokens generate the tools that apply themShipped

Hardware

  • FEARNOT Watch: firmware for a concert wearable worn upside-down on a lanyard: a pet, a countdown, a beacon, a setlist with haptics, a hype meter and more, every one checked on the deviceBuilt, verified on the device
  • CoreS2 exhibit: an LED signage frog that sleeps, wakes as you approach and blushes as a hand nears; the behaviour is kept separate from the hardware so it can be tested on a deskShipped
  • FAMILIAR: a generative creature on a palm-sized screen: a body that moves like fluid, senses that react to the room, a voice, and a listening modeBuilt, running on the device
  • Room-sensing bench programme: sensors for temperature, humidity, leaks and air quality across a set of buildings, reporting to a private backend with no vendor cloud; the most useful result was a decision not to build the room sensor in-houseBench complete
  • Core2 monolith, field-kit firmware, sensor-node enclosure, noise-sensor evaluationBuilt or documented

Strategy and planning

  • Trellis: a design toolchain for buildings: the regulations pinned to the clauses that matter, a floor-plan generator, and checkers that tell you what complies; applied to a real conversionFirst version
  • Four business plans: four independent plans on one matrix, each worked through to a decision with every claim marked as researched, analysed, assumed or validatedComplete
  • Financial model engine, 3D capability pack, agent harnessPaused or in design

Gates and tools

  • Citation census: refuses to publish a number that has no sourceRuns at build
  • The em dash guard: keeps one house rule true across a whole website, foreverRuns at commit
  • The system map: an index of every component, checked on every commit so it cannot drift from the codeRuns at commit
  • Drift-fix discipline, decision and changelog numberingRuns before commit

Operations

  • Company operations rebuild: every procedure in the company written down in one format, the shared drive reorganised so things can be found, and a training package for the teamComplete
  • Workforce analysis and hiring: an analysis of staff movement whose first conclusion was tested and overturned, and a hiring round screened to a shortlistDelivered
  • Knowledge hub, plain-English pass, statutory audit workstream, cloud cost audit, marketing dashboard audit, marketing data connectorsLive, shipped or active

How it was built

The projects vary. The way of working does not.

  • Review as a gateBefore anything ships, independent reviewers are asked to find what is wrong with it, not to confirm it is fine.
  • Verify on the real thingA clean build is not proof. Firmware is checked on the device; a deployment is checked by reading the live value back.
  • Observe before encodingAny assumption about how a folder, a file or a system behaves is checked against the real one before it is written into code.
  • Prove like-for-likeBefore trusting a comparison, run it on identical inputs first. If everything mismatches, the comparison is wrong, not the data.
  • Test the number, not the cellChecking that a cell exists proves nothing. The only test that means anything is on the value it holds.
  • Write findings down so they outlive the sessionEach programme keeps a running record of what was found, with corrections recorded as corrections.

Get in touch

Stuck on something?

Tell me what it is. I read every message myself and reply with what I would do about it.

It goes to my inbox and nowhere else. I read everything and reply myself.