Australian software engineer in London. Co-founder and CTO of Flowstate, putting human and AI labour on one ledger. I build products, cook, and drink far too much wine. Sometimes I write about AI. Sometimes I ski.
Open-weight models are now good enough to make self-hosting a financial question rather than a capability one. I modelled four options, an office rack, colocation, rented GPUs and per-token APIs, to find where each one actually wins. The answer turns on utilisation, and it isn't the one the internet keeps giving.
The capability gap between closed and open AI models, collapsing. Real LMArena data, told as a city.
Git worktrees only clone your code, so every worktree uses the same local database. One branch's migration breaks the rest. A schema change breaks the APIs. wtdb is a small Go tool that gives each worktree its own cloned Postgres database, env file, and node_modules.
The US government just ordered Anthropic to switch off two of its best models for everyone, citing national security over a jailbreak the company disputes.
Anthropic shipped Claude Fable 5 at $10/$50 per million tokens, a full tier above Opus and exactly double its price. It's a flagship, not a demonstrator, and it's the cleanest proof yet that AI pricing went K-shaped: the floor under Opus was set so the ceiling could be raised.
Your team is paying flagship AI prices for everyday tasks. Here's how routing and attribution turn a $3M undifferentiated OpEx bill into an amortizable asset.
Apple introduced Siri AI for iOS 27 but not in the EU, blaming the Digital Markets Act, while giving EU users the same assistant on the Mac. The line tracks which platforms the law designates, not any real difference in risk.
Most teams overpay for AI twice without noticing: they run everything through the default flagship model, and they buy the tokens on a metered plan they can't see into. An intelligent proxy fixes both. Here is the maths, a calculator you can drag, and where Flowstate fits.
Microsoft moved the menu again. A fast, funny tour of why integration beats flexibility, and why a company that stares at its rival instead of its user keeps losing the one person who matters.
How posts on this site actually get made — from a thought to a published page, and the per-sentence commenting layer I use to chase down where a draft is breaking. The comment box is open on this one. Try it.
A new paper from Stanford, Michigan, DeepMind and All Hands is the first open empirical study of how AI agents actually spend tokens at scale. The findings line up closely with what we are seeing at Flowstate, from a different angle.
AI is fantastic for extending thinking. The trouble is that the people crossing the line from extension to substitution cannot tell, from the inside, that they have. Including, possibly, the person who wrote this.
A tribute to Peter Whitehouse, my high-school IT teacher in Brisbane, who taught me memory management at fifteen, ran the IT club out of his spare time, and quietly built the architecture-thinking I lean on every day. Retired in 2022. Needed now more than ever.
The browser wars settled into a three-horse race. The AI race looks suspiciously similar — three players, scrappy upstarts, and one kid in the corner sucking on a jar of glue.
In 2012 Scott Weiss wrote that the best preparation for becoming a founder is working at a startup. Fourteen years and several failures later, here are the seven lessons I wish I'd actually listened to — on people, product, ownership, failure, and knowing when to step away from the keyboard.
Postgres' WindowAgg node doesn't propagate cardinality estimates. That single bad estimate can poison every downstream join decision. Sometimes the pragmatic fix is the right one.
Every discipline starts the same way. People do the work for years before anyone thinks to define it. Workforce Engineering is the discipline of deliberately designing, measuring and optimising how an organisation deploys its labour to produce outcomes.
Workforce Engineering is the practice of deliberately designing, measuring and optimising how an organisation deploys its labour, human and AI, to produce outcomes. Here's the framework, the six practices, and how to get started.
I needed to store massive OpenTelemetry data for AI spend analysis. Instead of reaching for BigQuery or Snowflake, I pushed 67,000 telemetry events per second into Postgres on a MacBook Air, while running analytical queries against it, to prove the boring choice scales.
We are currently adjusting our EBITDA to account for a sophisticated parrot that read the internet. Gartner says $2.52 trillion in global AI spend this year — yet only 14% of CFOs report clear ROI. Sequoia's 6:1 services-to-software ratio means the problem is about to get six times worse.
AI agents won't hollow out systems of record. But the platforms that fight agent access with rate limits instead of embracing structured protocols like MCP will lose to the ones that don't.
We have a digital identity problem. Not because verification technology can't inherently be private, it can. But, because governments are asking the wrong question. Governments ask "how do we identify people?" when what they should be asking is "how do we verify claims without identifying anyone?"
After years with Fastmail, we switched to Microsoft 365. Spam filtering, EU data residency, and ecosystem integration made the difference.
The EU sees open source as the path to digital sovereignty. They're right. But adoption without sustainable funding just means Big Tech keeps extracting value from European developers.
Why fluid planning beats rigid quarters, and how the best CTOs build cultures that adapt without chaos.
AI coding tools have crossed a critical adoption threshold. But the real shift isn't individual productivity—it's how engineering organisations allocate capacity across priorities.
On the case for moving away from anonymous "ghost" accounts toward a pragmatic internet—one where we can protect the vulnerable without silencing the dissident.
Exploring historic AI model prices, and Google's new Gemini 3 Flash at $0.50/$3.00 per million tokens—a 67% increase over 2.5 Flash.
It's IEEE Spectrum time again, so it's time to talk about why adding management layers early is the real problem behind $2 trillion in annual software failures. From Canada's Phoenix payroll system to the UK Post Office's Horizon fiasco.
Google's AI Overview recently told users that astronauts met cats on the moon. It suggested adding glue to pizza to stop cheese sliding off. For a brief period in early 2025, it claimed we were still in 2024. These aren't edge cases—they're symptoms of a much larger problem that's quietly dismantling the internet's economic foundation. LLMs are fundamentally breaking the internet's business model.
So you're spending 20 million on engineering. What are you getting for it? Software engineering organisations struggle to answer basic questions about costs, and PE scrutiny is changing the game. It's more important now than ever to treat engineering like the investment it is.
89% of malicious emails bypass SPF, DKIM and DMARC. Your inbox is already compromised—authentication protocols failed and nobody told you. Here's how AI catches what traditional filters miss.
Reflections on building Jamie, an AI assistant for email, and the decision to wind down Pragmatic after not finding product-market fit.
Analytics is curiosity about what works. Tracking is surveillance. I run Umami on Vercel with Neon Postgres for zero-cost, banner-free analytics. The irony? My joke LLM needs consent, but tracking 16,000 visitors doesn't.
The tech industry's rush to embed AI into every product is creating exhausted users and half-baked features. Real AI success comes from solving specific problems, not sprinkling chatbots everywhere.
AI is cheap for end-users right now, but only because venture capital is picking up the tab. The reality is that training and inference are still expensive, and someone has to pay. Billions raised by OpenAI, Anthropic, Mistral and others are spent on compute, staff and Nvidia GPUs. Consumers see subsidised pricing, but the underlying costs are enormous.
My take on the next version of the web given the failures of Web 3.0 and the need for a more organised, user-centric approach.
Reflections on meeting a team of scrappy, multi-faceted engineers and what makes the best startups work
A tale of 6,000 f-bombs and one accidental victory. What started as a simple integration for Jamie turned into a deep dive through WhatsApp's encryption, the Signal Protocol, and eventually finding a simpler path through DOM scraping.
What if your AI assistant could actually feel what kind of day you're having? Not in some creepy sci-fi way, but in a structured, measurable way that makes it respond like a human would—with actual understanding.
How switching from a broken WordPress site to Luma transformed It's Culture Innit's event management, proving that buying beats building when software isn't your core value.
Essential principles for defining Generic Business Objects (GBOs) in your system architecture to avoid expensive re-engineering and maintain clean data models.
A founder's warning about why early-stage startups should ignore cloud vendor sales calls and keep infrastructure simple to preserve runway and focus on shipping product.
How unnecessary CAPTCHAs, cookie prompts, and popup modals are driving customers away from websites, with practical advice on reducing friction.
Why VCs should back focused tools that integrate well with existing platforms instead of chasing the next monolithic unicorn that tries to do everything.
Why product teams need to embrace simplicity over complexity, focusing on clear blueprints and agility rather than endless documentation and rigid processes.
A detailed comparison of UK pensions versus Australian superannuation, highlighting how Australia's mandatory 11% employer contributions, lower fees, and better returns outshine the UK's restrictive system.
A reflection on how web development has evolved from the simple FTP upload days to today's complex ecosystem of frameworks, security concerns, and over-engineering.
A review of BunnyCDN as a European alternative to Cloudflare, highlighting its strengths in Edge Rules and constraint-driven development approach.
A detailed investigation into iCloud Mail's filtering failures, broken rules system, and unresponsive support that led to missed contracts and a migration to Fastmail.