All Articles
-
The Platform Under Your AI
Cloud-native produced a shared substrate because every application team was rebuilding the same primitives badly. The same consolidation is starting for AI: the gateway, the evaluation harness, the retrieval plumbing, the prompt registry, the routing controls. Which of those belong to a platform team, which belong to the feature teams, and the rule that separates them.
Read more → -
Someone Has to Own Evaluation
Cloud-native forced DevOps into existence because the old split between building and running stopped working. AI-native opens three ownership gaps: evaluation, the context supply chain, and the routing and cost decisions. A new title is usually how an organization avoids a change rather than makes one. Here is who should own each gap, and the one role that might actually be new.
Read more → -
You Don't Have to Rewrite Everything
The cloud era had the six R's, and they mattered because they gave teams permission to leave most of the estate alone. AI-native needs the same thing more, because the pressure to declare everything AI-native comes from outside engineering. Four criteria that sort the estate, and three answers instead of one.
Read more → -
Bolted On, Then Bolted Together
The distributed monolith was the cloud-native failure that looked like success. AI already has its own catalog: the load-bearing prompt, the eval that tests the model instead of the system, RAG on a data model never meant to be retrieved from, the agent with tool access, and the special-case integration. Each one, and the check that finds it.
Read more → -
You're at Level 2 and You Think You're at Level 4
Five levels, from "we added a chat feature" to "this is how we build now," each with an observable test instead of an adjective. Most organizations are two levels below where they believe they are, for the same reason they were on cloud maturity. The demo works, so the architecture must be fine.
Read more → -
AI Didn't Break Your Risk Matrix. It Broke Your Scores.
AI components belong on the same risk matrix as everything else. What changes is how you score them: a likelihood rubric built on how often a person meets the failure, and a severity rubric built from reach, reversibility, and detection lag.
Read more → -
You Didn't Change Anything. It Changed Anyway.
Services version on your schedule. Models drift, get deprecated, and improve on someone else's. Designing the boundary around a component you do not control, and the honest math on how much abstraction is worth paying for.
Read more → -
The Model Isn't Wrong. Your Context Is.
Retrieval, freshness, and data boundaries stop being back-office plumbing the moment a model depends on them. The context supply chain deserves the same rigor we learned to give service contracts, and it is where AI security actually lives.
Read more → -
Your Function Call Was Free. This One Isn't.
Distributed systems changed architecture because the network made a function call expensive. AI does it again, except the cost is money as well as latency. Model routing is an architectural decision, not an optimization you do later.
Read more → -
It Passed the Test. That Doesn't Mean It Works.
Every testing, monitoring, and reliability practice you own rests on one assumption: same input, same output. An LLM in the call path removes it. What replaces it is evaluation, thresholds, and verification as an explicit architectural layer with a cost and an owner.
Read more → -
When Everything Is Critical, Nothing Is
Reliability isn't only a technical property, it's an ownership property. Two decisions sit underneath every reliable system, what actually matters and who is accountable for keeping it up, and most organizations have made neither.
Read more → -
What the EU AI Act Actually Requires (And What It Doesn't)
The EU AI Act's reach extends to any company whose AI system's output is used by people in the EU, no European office required. A plain-language breakdown of the four risk tiers, what high-risk systems (hiring, credit, biometrics) must actually do, and what the Act does not require.
Read more →