AI Engineer Dojo Contents
Chapter 2

What a Healthy Eval Setup Looks Like

You don't need to build the system — but you should recognize one that's working. A healthy setup has two halves: a gate before you ship, and a watch after.

Offline evals run before release, on a fixed set of cases, answering "is this change good enough to ship?" Online evals run on live traffic, answering "is it actually working out there?" You need both: offline covers the cases you anticipated; online catches the ones you didn't. Underneath, every eval is the same four parts — a dataset (the cases), a task (running the system on them), a scorer (turning each output into a number), and a report (rolling it up, broken out by segment).

What good looks like Changes are gated by an offline eval in CI (quality can't silently drop), and a dashboard tracks live quality. The team can name their dataset, scorer, and how results are sliced.
Red flags "We test by hand before each release." No automated gate. No production monitoring. Quality is whatever the last person to look at it felt.
Decision Lab

Two teams, same model

Team A ships when "it looks good to the engineer." Team B has an automated eval that blocks a release if quality drops below a baseline, plus a live quality dashboard. Both have shipped for six months. Which team's roadmap can you actually trust, and why?

How to think about it

Team B — and it's not close. Team A's quality is invisible: it could be drifting down right now and no one would know until customers complain. Their velocity is fake, because every change is an unmeasured risk. Team B can move fast safely: the gate means a bad change is caught before users see it, and the dashboard means regressions surface in hours, not quarters. As a leader, fund the gate and the dashboard early — they're not overhead, they're what makes the rest of the roadmap believable.

Case study

HelpStack: three weeks of silent decay

HelpStack had a pre-release check but no live monitoring. A routine update quietly broke part of their system, and quality dropped for a whole category of questions — nothing errored, the assistant just started answering worse. With no dashboard watching live quality, the regression ran for three weeks before a support manager spotted the pattern in escalations.

The post-mortem was blunt: their offline check hadn't covered the affected cases, so it passed clean the entire time. They added a live dashboard tracking quality by segment; the next such break was caught in hours. The lesson for a leader: the offline gate proves a change is good on the cases you thought of, and only online monitoring catches the ones you didn't — you need both, and funding the dashboard is not optional overhead.

Running case · Meridian × Remi

This chapter: before launch, Meridian funds two things the CEO first calls "nice to have" — an automated check that blocks any Remi release if quality drops below a baseline, and a live dashboard showing answered / abstained / wrong-action rates. Both earn their keep inside the first month, when a well-meaning change would otherwise have shipped a regression.

Quiz · Chapter 2

  1. A healthy eval setup has:
  2. The four parts of any eval are:
  3. "We test by hand before each release" is:
  4. The reason to fund evals early is:
← Back Continue →

Evaluating Your AI · AI Engineer Dojo · aiengineerdojo.com