
AI that does the work. You keep control.
A Sarawak-based AI and software practice. We build the workflows, software, and content systems that carry real work — operated under human-in-loop authorisation, with the discipline of a security practice, for Malaysian enterprises, GLCs, and government.
Systems that carry live work — not pilots, not demos.
Axiom Blueprint is a small practice that builds AI into the work organisations actually do — the onboarding chain, the invoice run, the reporting pack, the content calendar. Systems that carry live work, that a team can operate without us, and that fail safely when they fail.
We have spent years operating and assessing systems that other people depend on, in regulated environments where a failure is not a rounding error. That experience shapes how we deploy AI: we have seen what happens when software is trusted further than it has earned. So every agent we ship is scoped, evaluated against real cases, logged, and reversible — and no action that touches a real system happens without structured authorisation.
Claude is the foundation. Axiom is the practice. The discipline is the deliverable — every time.
The AI does the work. A person signs it off.
A brief goes in. Agents do the reading, the drafting, the repetitive execution — scoped, evaluated, and reversible. But nothing touches a real system without a human authorising it, and every step is logged. What you get back is the work, done — with a record you can audit.
HOW THE WORK RUNS
One manager agent runs the projects. A person makes the calls.
Scroll through the floor, or pick a card. Change the time of day to see who is in. Agents never fake documents or move money without a person.

MANAGER → PROJECT AGENTS
The owner talks to one manager agent. It hands the work to project agents, one per project, and to an assistant for each team member.
> manager → build: slice C
BUILD → REVIEWERS
Every build goes to three separate reviewers: security, deployment and tests. Nothing reaches production until all three pass.
> reviewers: 3 of 3 passed
OWNER · MANAGER
The manager approves agreed work after the checks. Anything new or irreversible goes to the owner, who gets a digest twice a day.
> digest sent 18:00 · 1 decision waiting
THE RELEASE
The way back is planned before anything goes out. Each release is checked, released and verified, and rolled back if the check fails.
> verified · rollback ready
HOW WE KEEP AI UNDER CONTROL
The checks are part of how we work.
Four layers sit under the work. Pick one to see what it covers.
WHO ASKED, WHAT WAS DONE
Each job is written down: who asked for it, what was done and what was checked.
WHAT IT COVERS
- Who asked and what was done
- Review results kept with each release
- Full change history kept
NOTHING GOES LIVE UNREVIEWED
Three separate reviews must pass on the exact version. Then the manager approves agreed work; the owner decides anything new or irreversible.
WHAT IT COVERS
- Security, operations and testing reviews
- Reviewers never check their own work
- The approver is recorded with the release
CHECKED AGAINST THE PDPA
Personal data is identified before the build starts, and checked against the PDPA before launch.
WHAT IT COVERS
- Personal data mapped up front
- PDPA check before launch
A FAILED CHECK STOPS THE RELEASE
Every release is tested, and the real page is checked in a browser before it goes out.
WHAT IT COVERS
- Automated tests on every release
- A browser check of the real page
- Failed checks block the release
What we do.
We also run a cybersecurity practice — penetration testing, security training, and continuous monitoring. See the security work →
“Every action that touches a real system requires structured authorisation. Every output carries an evidence chain.”
- Sdn. Bhd. registered in Malaysia
- 200601010198 (729948-V)
- Built on Anthropic's Claude models
Brief us on scope. We respond personally within two business days.