This is a /now page.
The question I keep running into: organisations commission AI risk assessments and then do nothing with the findings. Not because the findings are wrong — because they're too technical to act on. That gap is what most of the work below is trying to close, in different registers.
Focus
Multi-agent AI safety and adversarial evaluation. The Failure First research keeps expanding; the last figures I published for it — 257 models, 142k prompts, 346 attack techniques — were current as of May 2026 and the recount is overdue. Alongside it, a run of work on what happens when you hand agents a whole production pipeline rather than a single task: The Ungovernable Body is an agent-operable film studio in a git repo, written up as a two-part production diary starting with whether that can work without flattening the politics.
Building
- Client projects — practical digital infrastructure for businesses that need it built properly and maintained by the person who built it
- This site — weekly sprints, shipping in public, with automated social publishing across Facebook, Instagram, Bluesky, and Twitter
- SPARK — a non-coercive AI companion for neurodivergent children, now with three distinct voices
- Afterwords — local voice output for Claude Code with voice cloning on Apple Silicon
- PAOS — a local-first agentic OS that runs on your hardware
- Dead Air — a voice-agent failure eval harness that models endpointing, barge-in, latency and mishearing before production does
- ClawdCraft — Claude Code living in a Minecraft server as a crab-skinned allay kids talk to in chat, with the boundaries that matter enforced in code
- Understory — a proposal-rigor lab stress-testing stacked low-capital income streams on a Tasmanian conservation parcel
- Homelab — Frigate NVR with Hailo-8L AI detection on a Pi 5, Home Assistant automations, self-hosted infrastructure
Writing
Publishing steadily — recent pieces cover compute and governance, safety benchmark transparency, AI alignment limits, community ownership models, robot dogs as security risk, and Perth's electronic music underground. Also running a NotebookLM pipeline that generates audio overviews and infographics for every post and project. See what's new for everything in one feed.
Available for
Two kinds of work: practical implementation for businesses that need reliable digital infrastructure without agency overhead; and specialist advisory for organisations navigating AI risk, governance, or security evaluation. Both start the same way — written scope, agreed price, deposit before work proceeds. Open to consulting engagements and conversations that start with a real problem, not a brief.
Location
Tasmania, Australia.
This Site
122,622 words written across 98 posts.
Recent Activity
Live from GitHub.