I built the first one because chat kept forgetting.

What began as a way to hold the context behind my own work became a client installation in Dubai and the architecture for a separate multi-agent system.

Type
Applied AI systems
My role
System architecture & build
Scope
Own system + 2 engagements
Since
May 2026
AI operating system connected to schedules, tasks, communication, research and recurring workflows

The useful part isn't the model. It's what the system remembers.

I was running a job search, independent client work and several side projects at once. A normal chat could help with one task, but it could not hold the decisions, priorities and context behind all of them. I kept having to explain the work again.

So I built a working system around the model: one home for each kind of information, a session history, decision and priority logs, reusable skills, deterministic tools and two specialist agents. I operate it myself, which means its weak points show up in real work rather than in a demo.

Once that foundation worked, I adapted the idea for two very different engagements. One was a full installation on a Dubai founder's own machine. In the other, I worked as architect and validator for a sequential-agent system that another team would implement.

11active skills
24Python tools
6repeatable workflows
~1 weekto install the client system

The three builds are at different stages: I actively operate my own system; the Dubai installation was delivered into the client's hands; and the separate multi-agent architecture reached its first three agents. I have not listed usage results for that last system because I do not have current operational data.

Four decisions behind the system.

Structure give each fact one home Templatise turn repeated work into skills Automate use code when rules matter Review audit what drifts AI operating system

The role changed with the problem.

I did not copy the same assistant three times. Each engagement needed a different level of ownership, architecture and handover.

Build · operate · improve
My own operating system

I designed and run the full system: structured context, decision history, live priorities, custom skills, specialist agents and deterministic tools. Four structural audits have helped me catch drift and improve how the system is organised.

Install · personalise · hand over
For a Dubai wellness founder

I installed a personalised assistant on her own machine in about a week, with an opening countdown dashboard, decision-tracking and scouting skills, five handover prompts and a way to diagnose file-sync conflicts herself.

Architect · review · validate
For a real-estate and dog-training business

I reviewed and approved the design of a sequential-agent architecture for another team to implement. The first three agents were delivered and diagrammed; my role was the architecture and validation, not the ongoing operation.

Under the hood, where it matters.

ClaudeClaude CodeMCP connectorsGmailGoogle CalendarGoogle DrivePersistent memoryCustom skillsAutomated workflowsDecision logStructured knowledge baseBackground tasks

Need someone who can turn AI into a working system?

I can help decide what is worth automating, build it and make it usable by the people who have to run it.

Let's talk →