Topic

Artificial Intelligence

Everything we publish on building with AI, from retrieval and agents to the evaluation practice that decides whether any of it can be trusted in production.

How we approach Artificial Intelligence

The model is rarely the reason it failed

Most AI work that stalls does not stall on model quality. It stalls because nobody agreed what a correct answer looks like, so there is no way to tell whether the last change helped. Evaluation is the first thing cut for time and the one thing that decides whether the system can be improved at all. We build the evaluation set before the pipeline, because a system you cannot measure is a system you can only have opinions about.

Retrieval is a data problem wearing an AI costume

Retrieval augmented generation fails on chunking, freshness, access control and duplicate documents far more often than on the model reading them. If two versions of a policy are in the index, the answer is a coin toss, and no amount of prompt work fixes it. The work is nearly always upstream, in the content and its permissions, which is unglamorous and is where the result actually comes from.

Agents multiply the blast radius, not just the capability

The moment a system can act rather than answer, every weakness it had becomes a change to something real. That shifts the question from accuracy to authority: what is it allowed to do, what needs a human, and what is the reversal path when it is wrong. We would rather ship a narrow agent with a hard boundary than a general one with a warning in the documentation.

Every page above is written to be used rather than skimmed, and each links back here and across to the others. Nothing on this page exists only to hold a keyword.

Other topics: Cloud, Cybersecurity, Enterprise Infrastructure, Oracle, Data and Analytics.