Artificial Intelligence
AI Evaluation & Quality
The largest gap between what is demanded and what is studied. Read this section even if you are targeting another archetype.
What the job actually is
Deciding whether a non-deterministic system got better or worse, with enough rigour that a team can ship on the answer. Golden sets, regression suites, human labelling protocols, LLM-as-judge and its many biases, and the statistics of drawing conclusions from a few hundred examples. It is measurement science applied to a system that gives a different answer every time you ask.
The arbitrage is stark: 59% of job descriptions name evaluation, and under 2% have it as a job title. Almost nobody prepares for it, and it is assessed in every applied-track loop. Getting good at this raises your performance in archetypes 1, 2 and 3 simultaneously — which makes it the highest leverage thing on this page.
Titles this hides behind
Prepare like this
- Build a real eval harness for something you have shipped. A golden set, a scoring function, a regression run, a report. Fifty well-chosen examples beat five thousand scraped ones.
- Learn the LLM-as-judge failure modes by name: position bias, verbosity bias, self-preference, and score compression at the top of the scale. Then learn the mitigations — swapping order, pairwise comparison, calibrating against human labels.
- Always carry a control. A number without a baseline is not a result. If your agent resolves 71% of tickets, the interview question is 71% against what — the previous version, a human, or a trivial keyword rule that would have got 65%. This single habit distinguishes strong candidates more reliably than any technical depth.
- Know your statistical power. On 200 examples, a 3-point difference is usually noise. Be able to say so, with the arithmetic.
- Design a human labelling protocol including the disagreement rate you would accept and what you do when annotators disagree.
Be ready for
- "How do you evaluate something with no single correct answer?" Rubrics, pairwise preference, and task-completion proxies — plus honesty about what each misses.
- "Your LLM judge agrees with humans 85% of the time. Is that good enough to ship on?" Depends entirely on where the 15% falls. Strong answers ask about the error distribution.
- "How do you stop your golden set from going stale?"
- "You improved the eval score and users complained more. Explain." The eval measured the wrong thing — and whether you can say that plainly is the test.
Compensation
A day that is actually representative
| Time | What you are actually doing |
|---|---|
| 09:00 | Triage the regression run. Four cases moved; two are genuine and two are the judge being inconsistent, and separating those is the skill. |
| 10:30 | Add cases. A support escalation revealed a failure class nobody had tested; you turn it into five golden-set cases before the fix ships. |
| 13:00 | Calibrate the judge. Sampling fifty cases against human labels and computing agreement, because a judge nobody has checked is an opinion generator. |
| 15:00 | Say no. A team wants to ship on a three-point improvement over 200 cases. You explain the noise floor, and this is the part of the job that requires standing. |
| 16:30 | Build. Automating the run so it happens on every change rather than when someone remembers. |
Real cases
Repeatedly, models have scored highly on public benchmarks partly because those benchmarks leaked into training data, which is why held-out and freshly-authored sets matter.
Use it to say: “It is why I insist on a private eval set and on checking dates. A benchmark everyone reports on is a benchmark everyone has trained on.”
The company was liable for what its assistant said.
Use it to say: “It reframes evals from an engineering nicety to evidence. If we cannot show what quality bar we shipped at, we have no answer.”
A three-week plan
Week 1 — build the harness
Days 1–2: Fifty reference cases by hand for something you have built, covering the failure modes you have actually seen.
Days 3–5: Scoring, baseline recorded, automated run. Then change one thing and measure the delta against the noise floor.
Week 2 — the judge
Days 1–3: Add an LLM judge and calibrate it against your own labels. Compute and report agreement rate.
Days 4–5: Demonstrate position bias and verbosity bias on your own data, then mitigate with pairwise comparison.
Week 3 — rehearse
Days 1–2: The multi-turn evaluation answer and the 85%-agreement answer, out loud.
Days 3–4: Write the one-page report: what you measured, against what baseline, with what confidence, and what you would decide.
Day 5: Five stories: an eval that caught a regression, a metric that measured the wrong thing, a judge you stopped trusting, a number you refused to ship on, a golden set that went stale.
The book for this field
LLM Systems Engineering
Seven chapters on the engineering underneath applied AI: the roofline and KV-cache arithmetic that explain every serving decision, quantisation with the evidence, serving stacks and the prompt-cache traps that cost real money, retrieval internals and whether long context replaced RAG, agents at the token level, evaluation with actual statistics, and what won in post-training.
Cross-cutting
Skills every archetype tests
These are shared across every field on this site — the same question asked in different vocabulary — so preparation here compounds rather than being spent once. The book above is specific to your field.
Practice the 22 AI questions → · Back to Artificial Intelligence →