Melbourne. I find the question hiding in a messy frontline problem, decide what "true" means, and build the system — people, data and gates — that keeps it true.
Until August 2026 I supervised loss prevention in a large retail store. Staff were binning loose price tags, and each one was a sign of possible theft with no time or place attached. That's where most of this started: turn a hunch from the floor into a data model, a workflow people will actually use, and a decision a human still owns.
I build by directing AI coding agents (Claude Code, Cursor, Codex). They write the code; I write the specs, the contracts they work under, the tests and gates, and I review the results. Every repo below says so.
| Repo | What it shows |
|---|---|
loose-tag-reporter |
Frontline evidence capture for a loss-prevention team: a stable location model, resend-safe offline sync, and an audited review workflow. Proof of concept. |
agent-operating-system |
How to run several AI agents on one repo: single-writer ownership, handoffs, delegation with veto and kill switch, scored critic seats, and build gates that stop claims drifting from the source of truth. |
evm-token-forensics |
Adversarial data forensics where every convenient data source can be gamed: "empty means unknown, never safe", simulation over trust, reading code instead of labels. |
preregistered-studies |
Two market-signal studies with hash-frozen hypotheses and a hold-out opened once. Both failed; both were abandoned. |
shadowbot-safety-controls |
A bounded-autonomy control plane for an automated agent: hard caps, 16 halt conditions, idempotent actions, human-signed promotion. Shadow mode only. |
sovereign-hive |
A private, Tailscale-only AI workstation with architecture decision records and an overnight builder/evaluator loop. |
tailscale-exit-node |
Infrastructure as code for a VPN exit node, with secrets kept off the command line and a real dry-run. |
hoopclaim |
A pickup-basketball app built around trust: location-verified check-ins, weighted peer ratings, unanimous results, and a 10-part security audit. |
- Find the question before the tool.
- Decide what "true" means, then make it a gate.
- Keep decisions with humans; make autonomy earn its scope.
- Write it down: rulings, handoffs and incidents become rules.
- Let the data say no — and set the threshold before looking.
- State status plainly: concept, proof of concept, live, failed.
Long term: education for South Sudan and its diaspora.