← BlogEngineeringMarch 18, 20266 min read

Evals before features: how we measure AI quality

If you cannot measure whether the AI got better, every prompt change is a gamble. Evaluation harnesses are the least glamorous, highest-leverage part of an AI codebase.

Traditional software has tests. Change the code, run the suite, know in minutes whether you broke something. AI systems lose that safety net the moment behavior depends on a model: outputs are probabilistic, "better" is fuzzy, and a prompt tweak that fixes one case quietly breaks five others.

The harness comes first

Before we build a feature on a model, we build the thing that judges it: a set of real inputs with reviewed expected outputs, scored automatically on every change. For extraction tasks the scoring is exact. For generation it is a rubric, model-assisted grading spot-checked by humans, and hard checks on the things that must never happen.

This sounds like overhead until the first regression it catches. A team with an eval harness ships prompt and model changes daily with confidence. A team without one ships monthly and argues about whether anything improved.

Tie it to the business metric

The last step is making the eval mean something commercially. Accuracy on a test set is a proxy; what we actually commit to is the business number it drives, claims processed per day, tickets resolved without escalation, search conversion. When the eval and the KPI move together, you can finally invest in AI the way you invest in anything else: with a number you can measure.

Want this working in your stack?