Most teams shipping AI features cannot answer the simplest question about them — is this actually good? Evaluation is not a testing chore. It is the definition of correct, and writing it down is the hardest and most valuable work in the project.
8 min readMost organisations have given their AI agents real power and borrowed credentials. The identity question sounds like plumbing, but it decides whether an agent is auditable, revocable, and safe to expand.
9 min readTraditional caches only recognise a question they’ve seen letter for letter. Semantic caching recognises meaning — and for the thousandth shopper asking ‘does this run small?’ in slightly different words, that difference is most of your bill.
3 min readAs teams wire up a dozen different models, a new piece of plumbing quietly becomes essential — the AI gateway that routes, governs, and pays the bills. Here’s the architecture, with an airline booking example.
5 min readThe clever one-liner prompt is fading. The real craft now is deciding what an AI model gets to see before it answers.
3 min readThe headlines chase ever-larger frontier models, but the more interesting move in 2026 is downward — small, capable models running on the device in your hand.
2 min readMost AI feature experiments share the same fate: impressive demo, uncomfortable production metrics, quiet retirement. Here’s the design thinking that changes the outcome.
2 min readEvery cloud migration case study ends with success. None of them tell you about the Thursday at 2am when the database connection pool hit zero during a live flash sale.
2 min readAfter two decades of inheriting other people’s burning platforms, I can tell you exactly when enterprise transformations die — and it’s not when the CTO finally admits it.
2 min read
