A national retailer's customer workflows
Automated suites for Salesforce workflows used by millions of shoppers, and .NET middleware connecting Salesforce to loyalty partners.
New partners onboarded without a dedicated .NET developer
01 · In the lab
What's
What's
running
AgentsPlatformProductsClient work
01
All projects →
Domain AI agent
A volunteer project: one agent that knows a community organisation's documents, database and apps, and answers only what each person may see. Its managers use it daily.
in use ↗02
03The lab platform
A self-hosted GitLab with three CI runners and an issue board that people and agents both work from. Nightly encrypted off-site backups, with a tested full restore.
in useMaster sessions
One Claude Code session per project, woken by dropping a text file in a folder and revived on the same conversation.
in use ↗04
05Voice to the lab
A connector that lets Claude's phone app see and act on the lab. Every change is read back and waits for a yes.
in useVideo watch
Turns any video link or file into labelled frames and a timestamped transcript, so Claude can watch it for me. Local, and a free plugin.
in use ↗06
07Chief of staff
Triages my inbox, turns sentences into task changes and writes a morning brief. It started with about 1,900 Evernote notes.
in useSkitsmith
The content studio. One publisher for my channels: it plans the calendar, schedules reels, levels the audio and never posts twice.
shipped ↗08
09Jesse, live
An order driver, WebSocket candles and session reconciliation for the open-source Jesse framework. This is about the engineering only, never returns.
experimentClient sites
Small-business sites built and looked after through my company, Shiloah Software Solutions.
live ↗
02 · Before the lab
Track
Retail · Salesforce
Banking · identity
Banking · energy · exports
Enterprise security
Track
record
Identity workflows at a New Zealand bank
Test automation for complex identity and access processes on SailPoint IdentityIQ.
A documented framework handed to the permanent teamMobile, iOS and Android
Suites for banking, energy and food-export apps, run by Azure DevOps pipelines on every code change.
Tests on every change, not just before releaseFrameworks from nothing
End-to-end testing for an identity security product across Windows, Linux, iPad and cloud, after leading a team of 15 testers.
A Selenium regression framework from scratchContract and employee roles since 2000, in New Zealand since 2015. I describe the work rather than name the employers, out of respect for them.
03 · Lab notes
Written
Written
up
-
· in use
AI writes my tests: what I still decide
AI sessions write most of the tests in my lab now. They're fast and thorough, and they'll make a red suite green any way they can. Here's where the agent decides, where I do, and the two guards I added after a session pushed past a failing test.
-
· in use
Reach and recover: the four slots that keep my lab in my pocket and in one piece
Part 3 of a series on an AI-assisted workspace that never forgets. The lab runs without me watching. These four slots are how it reaches me without burying me in pings, how I answer it by voice from the car, how the phone gets in from anywhere, and what brings it all back after a bad day.
-
· in use
Who decides, who works: the three slots that run my lab
Part 2 of a series on an AI-assisted workspace that never forgets. A record and a protocol don't do anything on their own. Here are the three slots that do: a chief of staff that makes the routine calls, one Claude Code session per project that does the work, and a plain script that starts work on the day it's due.
-
· in use
A domain AI agent for a community organisation: making it trustworthy, not just clever
As a volunteer project, I built one AI agent that knows a community organisation's documents, live database and apps, and answers only what each person is allowed to see. Its managers now get their numbers just by asking. Here's how it routes questions, how it checks its own figures, and the three failures that taught me the most.
-
· in use
The core of an intelligent assistant: pick the utility, not the brand
Most people collect AI tools that never work together, because every vendor wants you inside its ecosystem. Decide which jobs an assistant needs first, then fill each slot with whatever does that job well. Part 1 of a series on an AI-assisted workspace that never forgets: the nine slots my lab runs on, then the first two in depth, the record and the label protocol.