The lab
Most architects at this altitude design from reference diagrams. I run the silicon. This is where I load and quantise local models, mesh a hybrid platform end to end, and let my own agents run my own infrastructure, so that when I stand in a steering committee and say a design will hold, it is because I have already broken a smaller version of it myself. It is the bare-metal end of a boardroom-to-bare-metal job.
The silicon
I run an on-prem NVIDIA DGX alongside a cluster of Mac Studios. Not a rented endpoint, not a reference diagram: hardware in the rack that I load, run and break myself. It is where I pull local models down, run them, quantise them, and watch what the quantisation costs. When a model gets something wrong I want to know which of three things it was, the data, the quantisation, or the prompt, because the fix is different for each and guessing wastes a week.
So I build evaluation datasets for the workloads I care about and score against them rather than trusting a vibe. I experiment with fine-tuning local models on the DGX, mostly to understand the shape of the tradeoff, how much a small targeted model can close on a large general one before the cost stops being worth it. Owning the silicon is the point. It is the difference between designing an AI platform from a whitepaper and designing it having felt where these machines actually strain.
The platform
Around the accelerators sits the rest of a real platform. Proxmox clusters running LXC containers and VMs, container orchestration, Docker Compose for the things that do not need it, and workloads spread across multiple cloud regions. The whole estate is meshed and secured end to end with Tailscale, so a service on a Mac Studio at home and a node in a cloud region talk over one private fabric as if they shared a rack.
It has the unglamorous parts too, because those are the parts that fail in production: CI/CD that builds and deploys real services, monitoring and log aggregation and alerting, backup strategies I have tested by pulling the plug, and kill switches that cut spend when an agent goes rogue. I build the architecture, break it, fix it, and only then draw the diagram, because the diagram should document what survived, not what I hoped.
Retrieval and memory
Retrieval is easy to demo and hard to get right, so I run it over corpora that punish shortcuts. Capitalogic, my Australian tax-return drafter, retrieves over an ATO corpus with hybrid search: BM25 for the exact statutory phrasing that dense vectors blur, pgvector for the semantic neighbourhood, and a live fetch behind an SSRF guard for the things that change. Dense-only retrieval quietly loses the precise legal wording, which in a compliance domain is the whole game.
The other half is memory. I run a shared cross-agent vector store that multiple agents read and write, so a fact learned in one session survives into the next and across agents, rather than every run starting cold. It is retrieval-augmented memory as infrastructure, and it is what turns a set of one-shot agents into something that accumulates.
Orchestration
The lab is also where my agents live. A fleet of specialised agents, each with narrow tool access and current context, handles the operational work I used to do manually: cloud provisioning, identity, network configuration, monitoring, code delivery. An orchestrator routes a request to the right agents, runs high-stakes work through an evaluator loop before anything executes, and asks me through Telegram when a decision genuinely needs a human. It runs on the same platform I test client architecture on, which is the point: I am not demoing agents, I am running my own infrastructure with them.
What I design to
I am honest about the line between what I own and what I design to. The DGX and the Mac Studios are mine. Factory-scale silicon is not something anyone keeps at home, so NVIDIA HGX and GB200-class superclusters, and the AMD ROCm and Instinct MI300 to MI400 stack, I track and design to as reference architectures, and I stay vendor-neutral on purpose. The same goes for the fabric and storage plumbing of a supercluster, RoCEv2 and InfiniBand, NVMe-oF and GPUDirect Storage: design-authority level, validated at lab scale, sized for the real thing. The lab does not pretend to be a national platform. It makes me trustworthy when I design one.
The lab is the hands-on proof. The other half, the investment cases, the reference blueprints, the governance that decides whether an AI factory actually works, is told in the story. The builds it produces are in selected work.