AI agents in production
Multi-agent systems working continuously against live business systems, with real failure modes owned and fixed. Not demos.
Scale, in context
- 20 years
- Building and operating data, infrastructure, payment, and email platforms across New Zealand, Switzerland, and the US market. The judgement on this site comes from operating systems, not observing them.
- 270M
- US persons covered by the identity and intent platform I'm responsible for at Blackpearl Group: a measure of the platform's reach and of the governance obligations that come with it.
- 700B+
- Tokens through a single production model integration in one recent quarter. That is sustained production throughput on one of several products we ship, not a benchmark run.
Three things, one person
Data at national scale
Snowflake infrastructure processing intent signals across 270 million US persons, packaged into commercial data products.
Governance that holds
Author of Blackpearl Group's AI Agent IAM Policy, now in production, and led security remediation programmes across the portfolio.
Current focus
At Blackpearl Group I lead data engineering, analytics, information security, and TechOps across a distributed team. I stay close to the architecture: when there's a security gap I audit the codebase myself, and when we need a new pipeline I write the SQL.
Outside the day job I run my own agent infrastructure, and in August 2026 I organised New Zealand's first Claude Impact Lab, a civic hackathon with Wellington City Council.
Recent work
Delivered
Wellington Claude Impact Lab
104 applications, fifty builders, ten working prototypes for Wellington City Council's emergency management team, in one day.
Open source
@swampratnz/agent-base
The npm package extracted from my community-agent work, MIT licensed, public on npm.
In production
Conn, a multi-agent chief of staff
11 containers on a Mac Mini, Claude Agent SDK, credential proxy, egress control, and a self-improvement loop that cannot approve its own changes.
Writing
Essays on building and governing production AI systems.
- 2026-08-10
Production AI agents need identities, not shared credentials
An agent that acts on live systems is a workload, and it deserves the same identity discipline as any other workload. Shared credentials are how agent projects become incident reports.
- 2026-08-10
Governance is an enabler for agentic systems
The teams that treat governance as paperwork end up with agents nobody trusts. The controls are what let you extend autonomy instead of revoking it.
- 2026-08-10
Why technical leaders should stay close to implementation
I am a Chief Data Officer who still writes SQL and audits codebases. That is not nostalgia for the tools. It is how the strategy stays honest.
Start a conversation
Data, AI, and agent governance
Comparing notes on production agents, identity for autonomous systems, or data platforms at scale.
Get in touchSpeaking and events
Talks on agentic AI in production, AI governance for boards, and what the Wellington Claude Impact Lab taught us.
Invite meWellington and civic AI
The Claude community, meetups, and what AI can usefully do for public institutions in New Zealand.
Connect