Architecture
On-prem AI is not a cost decision. It's a sovereignty decision.
Why the on-prem-versus-cloud argument for regulated AI is usually framed around the wrong variable.
Architecture
HITL is not a checkbox — it's a write-path lock
Human-in-the-loop oversight only means something if it's enforced structurally, at the point where a system writes a decision, not in a UI someone can skip.
Architecture
Designing multi-agent systems that don't become multi-agent messes
Boundaries, ownership, and the specific failure modes that show up only once you have more than one agent.
Architecture
The Architecture Decision Record as the most underrated design artefact
Most teams write ADRs nobody reads. Here's why the format is right and the discipline around it is usually wrong.
Architecture
The token limit is not the bottleneck. Your memory architecture is.
A larger context window doesn't fix a system that was never designed to remember the right things.
Architecture
Zero-trust is not a network configuration — it's an architectural philosophy
Treating zero-trust as a VPC setting instead of a design principle is why it fails to actually contain anything.
Architecture
The €40M problem: why the most expensive engineering failures are invisible until they're not
A case for instrumenting the failure modes you haven't seen yet, not just the ones you have.
Architecture
The architecture of accountability: HITL, XAI, and the EU AI Act deadline
What Article 14 and Annex III actually require of a system's architecture — not as legal summary, but as design decisions.