Turning principles into working systems
At SmaariX, I build agentic systems that do useful work within explicit limits. They share one backbone, a platform Spine, and today it carries two products: a supply chain platform and a hiring workspace. In both, I’m clear about what an agent can propose, what it can act on, and where judgment stays with people.
Supply chain and hiring make those boundaries real: decisions move goods and affect people, and they need to stand up to scrutiny. Singapore’s Agentic AI Governance Framework reflects the same principle: greater autonomy still requires meaningful human accountability. IMDA ↗IMDAModel AI Governance Framework for Agentic AIView framework ↗
- One backbone
- Two products on the SmaariX Spine: supply chain and hiring
- Two proofs
- Connected, and seen running after deployment
- Found and fixed
- A gate that stayed green over a stopped mechanism
- In progress
- Cross-agent handoff, next to be proven
One backbone, two products
The SmaariX Platform Spine. Sixteen layers, from the screen to security and delivery, holding 179 researched capabilities. The layers stay the same across domains; what changes is the industry, the requirement and the decisions an application has to support. Each product takes what it needs through replaceable connections, so a component can change without rebuilding the system.
The supply chain platform. A supply chain platform I designed from years of logistics data-integration work, where the answer usually existed but sat scattered across screens, reports and people. Planners ask the Copilot about the screen they are on. Answers cite the figures in view and the agentic knowledge hub, refuse what the user isn’t cleared to see, and propose next steps a person approves. Every question first meets a local decision model that routes it to the fastest safe path: a prepared answer, a curated answer or full retrieval, with a second person approving every curated answer. Quality is measured against a graded set of questions: grounded answers, citations, correct refusals and speed. It runs today on representative data.
SmaariX.com, the hiring workspace. Job seekers, recruiters and hiring managers each see the context for their next decision, and a person stays accountable for it. Copilots work only with the candidate records they’re permitted to access and act within explicit authority. An action is either permitted or it isn’t; the boundary doesn’t quietly expand with use. It is built, and shown on request.
A durable memory plane. I use it every day; how it keeps knowledge apart from authority is on the Systems page.
How I verify it
Every mechanism needs two proofs: that it is connected to the live system, and that it actually ran after deployment. If either proof disappears, the build fails.
If a mechanism cannot prove itself, I don’t call it done. The same standard applies to reporting: figures come from source, but the inputs, logic and output still have to be checked.
We even found a gate staying green after the mechanism it watched had stopped firing. We fixed the proof, not the story around it.
This is the settlement-system instinct I bring to my own engineering. A claim without a record isn’t a claim.
Where it applies
Supply chain and hiring are where I’m applying the pattern, not where it ends.
Agriculture, renewable energy and pharmaceuticals have very different data, but the decision pattern can be the same: gather the evidence, surface the exception, keep judgment with the right person, and retain the record.
What changes is the evidence and who is qualified to decide. What stays constant is accountability for the outcome.
That decision boundary, the point where technology stops and human judgment begins, is the transferable part.
This is how I enter a new industry: I learn its language and its decisions from the people who make them, and connect them to a backbone already built to adapt.
Built to change
The agent runtime and models are deliberately replaceable. Changing the technology should mean changing a component, not rebuilding the platform.
That is not a bet on which tool will win. It is recognition that this field will keep changing.
The faster a technology changes, the easier I want it to be to replace.
In progress
Cross-agent handoff: passing work between agents without losing context or authority.
It is in progress. Until it can prove itself in practice, I won’t describe it as proven.