Current Build

What is being built now - a graph-native data core, a modular product on top of it, and the AI operating model alongside.

I'm currently running a deliberate architectural experiment in parallel with my full-time role.

The core

A graph-native, modular data architecture framework I'm building from scratch. Eight primitives, full mechanism set, comprehensive test suite. Two hard rules: the core never imports a product, and the core never uses new. Built to solve the framework drift problem most product suites suffer from - three years in, you're maintaining four codebases pretending to be one.

The product

A modular consumer product for business operators, built as the first real consumer of the core. The product isn't a demonstration of the core - it's the proof. If the framework can't carry a real product to production, it isn't a framework.

The methodology

I don't use AI as a productivity tool. I run it as a distributed engineering organisation - one stream for planning and specification, another for hole-poking and quality control, a third for test-as-spec - with myself as architect, reviewer, and integrator at the centre. Every member of that organisation has perfect recall, zero ego, and no architectural opinion of its own. The architect at the centre still owns every call.

This is not "AI built it." Two real framework bugs caught by the test suite so far. None by the AI. Both by the discipline. Even at the build level, the ratio holds: the AI is the 30%; the architecture, the discipline, and the judgment of what to ship and what to throw away - that's still the 70%, and that part is still mine.

How it started

The catalyst was a project delivered under political fire. In 2025 a client faced a forced choice: after an unauthorised third-party leads system created a compliance mess, they could adopt their global IT provider's ageing solution - and, one system at a time, surrender the digital autonomy we had spent years building together - or I could build them something better, fast. I read the stakes for what they were, agreed, and made a decision I had never made before: a deliberately compressed discovery, made transparently - my team knew the politics, knew why, and knew I would be available at any hour to close in conversation the gaps a fuller brief would have closed on paper. The system shipped and succeeded: the client's MDs have benchmarked it above the global provider's own solution, mandated it as the single source of record for all marketing and sales activity - "if it is not recorded there, it did not happen" - scheduled it for showcase to the visiting global principal in October 2026, and mandated its Malaysian rollout for Q1 2027. But the delivery friction taught me what the accolades did not: my ceiling as an architect was the fidelity with which what was in my head could be transferred. Things I had specified in full were still sometimes missed; things I compressed cost my team dearly. I wanted to know whether the losses in translation were a communication problem or a medium problem. So I ran the experiment properly - build a product end to end with AI as the development team, and be involved in every layer. Not to become a programmer, but to know enough, from data model to infrastructure, to push back.

The graph moment

The experiment produced the most consequential technical correction of my career. Describing a data model I had specified five years earlier for my first platform - entities carrying any number of tags, tags carrying properties and values, everything connectable to everything - an AI gave me the name I had never had for it: a graph data model. I verified the answer independently across multiple models, established that it could be implemented on standard MySQL at costs acceptable at every scale my instances actually reach, and confirmed it had been feasible all along. Then I rebuilt my product roadmap around it. The vision had been right for five years; what had been missing was the vocabulary. I do not intend to be missing vocabulary again.

The documentation

The corpus behind this is not described; it exists. The framework is specified in primitive-by-primitive documents in which every architectural choice is recorded as a decision block - what was decided, why, and what it supersedes. Independent hole-poker review rounds produced locked decisions marked as locked; open questions are listed as open. The first product carries a formal alignment document reconciling its design against the framework version it was authored against. The product specifications that preceded it - including the queue-management system now in client hands - separate deployment-specific configuration from core product behaviour marker by marker, because the difference between a project and a product is knowing which is which. The documentation itself runs on a living-documentation platform I built for the purpose, rendering one source of truth through four personas: an executive summary view, full developer depth, an operational how-to, and a verification view. Architecture documentation is available to hiring teams on request.

The operating arm

The same pattern - structure first, then work; judgment stays human - runs the operating business daily. Weekly invoice orders that once took hours of transcription take minutes, with tax-handling rules encoded and one step deliberately kept manual as the verification layer. The conglomerate group buy runs on per-unit utilisation tracking, differentiated invoicing modes, a forward-utilisation mechanism I designed, and renewal allocation models built from utilisation history. Legal governance runs through two purpose-built assistant personas - every clause interpreted, tested against policy, redlined by my own hand with AI never touching the document, every returned version checked for untracked changes (caught one) - a methodology disclosed to Group Legal from the start and validated by them as a legitimate first layer of defence - most recently in writing, after the Head of Legal reviewed a framework agreement I had negotiated end to end and cleared my positions on every substantive clause she had queried. Tribal knowledge from three chat working groups was extracted, cross-analysed, and codified into the content policies now attached to every insertion order. And one unused platform capability became a billed service run end to end by one person with AI (case study above). None of this replaced anyone. It let one person carry the working volume of a function while every decision stayed human. That is the 30/70 applied to AI itself: the model is the 30.