ExpertLesson 6 of 611 min read
Assessing what you build: the quality loop
A course from the Academy: Building with AI: the engineering level
Go to the quiz ↓Without measurement, improving an AI system is like cooking without tasting: every prompt or model change is a gamble. Serious teams distinguish themselves less by their models than by their evaluation loop.
Why public benchmarks aren't enough
MMLU, HumanEval and other rankings measure general capabilities, on questions that don't resemble yours (and that models may have encountered during training). A model can shine on benchmarks yet disappoint on your French maintenance contracts. The only metric that matters: performance on your use case.
The golden set
Put together 50 to 200 real, representative cases: the question asked, the context, and what a good answer must contain. Include easy cases, hard cases, known traps, and cases where the right answer is "I don't know." This set becomes your benchmark: every change is measured against it.
How to score answers
- Exact checks where possible: is the JSON valid, is the figure correct, does the citation exist? Reliable and free.
- Criteria rubric for the rest: accuracy, completeness, tone, absence of fabrication, scored separately rather than as one vague overall grade.
- LLM judge. A model scores answers against your rubric: it scales, but calibrate it first (have a human score 30 cases and the judge score the same, then compare) and watch out for its known biases, like preferring longer answers or the first position in a comparison.
The loop in practice
Every change (new prompt, new model, new RAG chunking) goes through the golden set before production: that's your regression test. A score that rises on hard cases but falls on easy ones is a signal, not a detail. In production, keep measuring: user feedback, a hand-reviewed sample each week, tracking refusals and errors.
You've finished the course. With the prompt ladder, RAG, fine-tuning, agents under guardrails, layered security and this quality loop, you have the full toolkit of modern AI engineering: the one that makes the difference between an impressive demo and a reliable system.
Key takeaways
- Public benchmarks don't measure your use case.
- Golden set: 50 to 200 real cases, the benchmark for any change.
- Exact checks first, criteria rubric next, calibrated LLM judge for scale.
- Every change goes through the golden set before production: that's your regression test.
Key concepts
Complete the lesson
A 4-question quiz: 3 correct answers are enough, plus an exercise to practise. It opens in the Academy, on this same lesson.
Take the quiz and complete the lesson Free. With an account, your progress is saved and leads to the course certificate.