Evaluation over gut feeling: how to measure the quality of an AI system
“Feels better” is not a metric. How to build a test set that surfaces regressions before customers report them.
Every change to a prompt is a deployment. Nobody notices because no test turns red. That is precisely why AI features can quietly degrade over months.
Build a test set from real cases
The best starting point is the support inbox and the logs. Forty to sixty real requests, a third of them deliberately difficult edge cases, are enough to begin. Invented examples test the author's imagination, not the system.
- Collect cases and record what a good result looks like for each one.
- Write the grading criteria before generating the first output.
- Run every version against the same set and version the results.
- Add new production failures as new cases — the set grows with the product.
Grade automatically, but carefully
Using a model as a judge scales well and is surprisingly reliable with clear criteria. It has one quirk: it favours long, confident-sounding answers. Optimise for that alone and you get systems that sound good and help less. A manually reviewed reference sample is part of the setup.
The real gain is cultural: once numbers are on the table, the debate about whose prompt feels better ends.
Author
Keep reading
Agents in the mid-market: what really happens after the pilot
Almost every company is running an AI pilot by now. Few make it into day-to-day operations. Why that is — and what the exceptions do differently.
Building a small RAG system: what you actually need
A RAG system sounds like heavy infrastructure. For a small project, four building blocks are enough: documents, an embedding model, a vector database, and a language model. Here's how they fit together conceptually, using Google Gemini and ChromaDB.
Creative production: how much does AI really take over?
Generated creatives have become cheap. That did not reduce the work — it moved it. A look at where the bottleneck sits now.