I run what I recommend. I have already built and broken the architecture I sign off, before I ask anyone else to trust it.
I'm Nav, a principal-level architect for AI factories and data centres. For twenty-five years I have decided where compute physically lives, from hazardous-area edge sites and liquid-cooled GPU farms to the AI infrastructure I design at national scale today. My work runs from boardroom to bare metal: I can defend a multi-year investment case to a steering committee and then go and run the silicon myself. During the day I do that for large regulated organisations. At night I build the tools I wish existed, on a DGX Spark and a Mac Studio cluster in my own lab. Everything I make, I use.
Building a multi-agent robotics setup where agents coordinate to gather and analyse real-time data. Working on an air-gapped home AI that controls the house without touching the cloud. Running, quantising and evaluating local models on a DGX Spark and a cluster of Mac Studios in the lab. Shipping OpenDray, a terminal cockpit for AI coding agents.
Local models on the DGX Spark, frontier models in the cloud, running as one system. Shared vector memory, a shared knowledge base, central secrets, round-table deliberation, Telegram and Slack as channels, one skill and plugin repository across every model, MCP, backups and high availability. Open source.
Read this post →Capitalogic drafts Australian personal tax returns. The design problem: tax is deterministic law wrapped in fuzzy paperwork, and a language model must never be trusted to do the arithmetic. So the maths lives outside the model entirely.
Containers, orchestration clusters, multiple cloud regions, a mesh network tying it all together. I run a scaled hybrid lab because I refuse to recommend architecture I haven't broken myself. If I'm going to stand in a steering committee and say something works, I want to have broken it first.
You need a kill switch that doesn't depend on you being awake.
go:embed bakes the Flutter web build into the Go binary. One file to copy, one process to run.
I added a Telegram bridge to OpenDray so I could check on Claude sessions from my phone. It was supposed to take a Saturday.
ShopSerp watches 250+ stores across 17 countries. It started because I was shopping for a camera lens.
I kept doing the same infrastructure tasks every week. So I built agents to do them instead. There are 14 of them now.
AI factories, GPU superclusters, high-performance storage, and the air-to-liquid transition. The current chapter, where the thermal, fabric and storage instincts from twenty-five years decide where AI compute physically lives, and where the investment case gets defended.
User provisioning, license assignment, mailbox permissions, security policies. Every tenant slightly different. Every change a potential landmine.
Landing zones, the 6 Rs, guardrails, and the compliance that a financial regulator actually checks. Years of principal work moving multi-environment enterprise platforms from self-hosted to public cloud, safely, at multi-million-dollar scale.
The hard part wasn't the automation. It was testing it. You can't simulate a real data center failure in a staging environment.
Every streaming platform stores metadata differently. Song titles don't match. Artist names have variations. ISRCs are unreliable.
Key management, transaction signing, cold storage. No established patterns existed. No frameworks.
End to end crypto mine farms: dense GPU clusters, immersion and direct-to-chip cooling, power distribution, and the low-latency pool that fed them work. The rigs are long gone. The craft is exactly what an AI factory needs now.
Refineries, power plants, mining platforms. Edge data centres in hazardous zones, where a stray spark is a legal event and the network cannot afford to hesitate. This is where the discipline started.
The architecture work, AI factories, data centres and the migrations behind them, is told in the story, and the hardware it runs on is in the lab. These are the builds at night: the tools I make because I want to use them, chosen here for the design problem rather than the demo.
An AI agent that drafts Australian tax returns. The hard part is that tax is deterministic law wrapped in fuzzy paperwork, so the maths lives outside the model entirely: a Decimal-only tax engine the model is never allowed to do arithmetic in, an agentic tool-use loop, document extraction via tool-use as typed structured output, and a BM25 + pgvector RAG layer over an ATO corpus with an SSRF-guarded live fetch. Compliance-grade rigour applied to a night project. Read how it is built →
A fleet of specialised agents that runs the operational work I used to do by hand: cloud provisioning, identity, network config, monitoring, code delivery. The design bet is that narrow, current context beats broad, stale context, so agents are stateless between calls and read live state every time. An orchestrator routes work, runs high-stakes changes through an evaluator loop, and asks me on Telegram only when a decision needs a human. Read the design →
A control plane that runs local and cloud models as one system: round-table deliberation across models, a shared vector memory and knowledge base, central secrets, one skill and plugin repository across the fleet, MCP integration, Telegram and Slack as channels, and backups with high availability. Built with the discipline of enterprise infrastructure and released open source. Read the design →
A self-hosted FastAPI render server on a Mac Studio running local generative models for music and motion video, reachable over a Tailscale mesh with a SOCKS handoff. It replaced a cloud rendering bill with hardware I already own. Local model ops at the edge, and a straightforward piece of cost engineering.
A shared semantic memory store that multiple agents read and write, so a fact learned in one session survives into the next and across agents instead of every run starting cold. Retrieval-augmented memory treated as infrastructure. It is what turns a set of one-shot agents into something that accumulates.
I was shopping for a camera lens and checking five different websites. That annoyed me enough to build an agentic crawler that watches 250+ stores across 17 countries, tracks price history, and tells me when the best deal appears.