Tech, teams, and the bits in between.

A deterministic layer that tracks what a coding agent has already read and changed lifts first-try success by 8 points while cutting cost by 29%.

Models find the right file for 92% of required deletions but cut the exact line under 52% of the time. Test for removal and pass rates fall from 63% to 42%.

An agent working on files and app state instead of screenshots scored higher at a ninth of the cost. Plus harness cost, hidden reasoning, and agent security.

A technical look at Ukraine’s drone stack, its battlefield engineering loop, and the evidence behind balloon-launched aircraft.

A parking camera joined two visits together, then handed Keith Miller the job of proving it wrong.

38,064 XML files, 241,456 amendment events, the oldest from 1872. What the entire NZ statute book looks like as data, and why counting 'a change' turned out to be the hard part.

I scraped all 41,301 Show HN posts from the last year. Half the comments a launch will ever get arrive inside 7.2 hours, and success doesn't slow the clock.

In 2025 Parliament sat 644 hours and spent a third of them in urgency. The churn isn't a personality flaw. We built the company this way.

Auckland Light Rail: six years, $228 million, zero metres of track. A running tally of what changing our minds costs.

I wired a live C. elegans connectome into the control loop of a Claude coding agent, then watched what the worm did when the code broke.

The Swiss cheese model trains you to count slices. The holes that actually kill you are the ones you never tested.

An interactive read-through of Robert Greene's 18 laws — 18 cards, full essays inside each.

How I build: small, narrow, niche, AI-native.

On hardening as a senior leader, the cost of walls that read as wisdom, and the quiet work that keeps a dent from becoming a crater.

Pocket shut down in July 2025. I needed somewhere to dump the articles I keep telling myself I'll read later. So I built a Chrome extension that does exactly that and nothing else. Local-only, no accounts, no tracking.

PMF gets sold as a milestone. It's actually a gauntlet that bends founders, breaks teams, and quietly poisons technical decisions. Most of the damage isn't from missing PMF — it's from how you behave while looking for it.

On the personal version of change management — the long, weird middle bit between who you were and who you're becoming.

ST-ResNet decomposes crime patterns into three temporal scales and models each one separately. Clever architecture, but does it actually help with only four years of NZ data?

You pay Big Tech about $1,000 a year in attention. Here's how to read the meter — and why building your own is suddenly cheaper than opting out.

ConvLSTM was built for weather radar. Turns out predicting crime on a grid is basically the same problem. Here's how it works and what it learned.

AI made software cheap to build but not cheap to scale. Evolutionary architecture with domain-aware AI agents is the missing piece.

Wiring up an MCP server so Claude Code can spawn isolated VMs, the SSE-to-Streamable-HTTP migration that broke everything, streaming architecture, and what productionising this would actually look like.

I built OpenHealth because Claude's and ChatGPT's Apple Health connectors are US-only. Drop your export.zip, get seven markdown files any LLM can read — browser, CLI, or phone-to-desktop over WebRTC.

Anthropic's Glasswing is for enterprises with six-figure security budgets. The rest of us need open-source security agents that learn our systems and act autonomously.

Building a Firecracker MicroVM rootfs, wiring up TAP networking, debugging UFW's FORWARD chain, and writing a guest agent with raw vsock syscalls in Go.

Before deep learning gets a chance, we need to know how well stupid-simple models perform. Turns out, they put up a real fight.

Anthropic has an internal platform for running Claude Code in full-permission isolation. I built my own version with Firecracker MicroVMs, Go, and a $400 Ryzen box.

Four architectural patterns from NanoClaw's tiny codebase that translate directly to production web apps and SaaS — credential sidecars, infrastructure-level isolation, polling loops, and Postgres as your message queue.

We're still shipping hand-crafted frontends while AI can generate entire interfaces. What if the browser itself generated the UI from an API manifest and your preferences?

Turn any Android phone into a programmable SMS gateway for your SaaS — no per-message fees, no carrier contracts, no vendor lock-in.

Before throwing deep learning at Auckland crime data, you need to actually look at it. Seasonal patterns, spatial hotspots, and the sparsity problem.

Crime records come with names and codes but no coordinates. Here's how we joined 1.15 million records to Stats NZ boundary files and gave every crime a place in physical space.

Turn a million geo-tagged crime records into a 4D tensor by overlaying a 500m grid on Auckland. Crime prediction becomes video prediction.

NZ Police's crime dataset is publicly available, but it's UTF-16 encoded, full of trailing periods, and 32% of records don't know what time the crime happened. Here's how we cleaned it up.

NZ Police publish over a million crime records openly. What happens when you point deep learning at them?

I pulled apart NanoClaw's 8,000-line codebase and found six architectural patterns that most SaaS teams should steal — credential proxies, container isolation as authorization, two-cursor message processing, and more.

The engineering career ladder was built for when code was expensive. Code is now cheap. What's valuable has fundamentally shifted.

AI coding agents burn too many tool calls just figuring out where things are. I built codemap to fix that — cheap model indexes, cheap model selects, big model codes.

Introducing c302 — a research project that uses the C. elegans roundworm's neural dynamics as a behavioural control layer for an LLM coding agent.