Insights
Field notes from production
Practical writing from our engineers and architects on what actually works in enterprise AI, cloud, data, Microsoft and security programmes — including the things that did not work.
Your RAG system is not failing on the model — it is failing on chunking
When a retrieval assistant gives confidently wrong answers, teams reach for a bigger model. The cause is almost always upstream: how documents were split, what metadata survived, and whether reranking exists at all.
Lift-and-shift is not a migration strategy
Relocating virtual machines to a hyperscaler usually raises your run cost. What actually delivers the business case is the modernisation you were hoping to defer.
When not to use Kubernetes
Kubernetes is excellent technology solving problems most organisations do not have yet. Here is the honest test for whether you are one of the ones that does.
Copilot will find everything you overshared
Microsoft 365 Copilot honours existing permissions perfectly. That is precisely the problem — it surfaces a decade of accumulated oversharing at conversational speed.
The rewrite you are planning will not finish
Big-bang rewrites of business-critical systems have a poor completion record, and the reasons are structural rather than technical. Incremental strangler migration is slower on paper and faster in practice.
Data contracts, minus the ceremony
Most data contract initiatives collapse under their own governance weight. A lightweight version — schema, freshness, ownership, enforced in CI — captures nearly all of the value.
Two hundred Power Apps and nobody knows who owns them
Low-code adoption succeeds, then becomes the problem. Environment strategy, DLP policy and solution ALM are what separate a productive maker community from unmanaged shadow IT.
A flaky test suite is worse than no suite
Once engineers learn that a red build might mean nothing, the suite stops providing information. Treat flakiness as a production defect with an owner and a deadline.
Nobody reads your dashboards, and your alerts wake people for nothing
Alert fatigue is not a discipline problem. It is a design problem: alerting on causes instead of symptoms, and on resource metrics instead of customer experience.
The penetration test report nobody acted on
A 200-page report with 340 findings produces paralysis, not remediation. What changes outcomes is exploitability context, business impact and a sequenced plan with named owners.
No articles under that topic yet.
Subscribe
Engineering notes, once a month
One email a month with what our teams learned in production — including the approaches that did not work. No product marketing.