AI Engineering

Write the evaluation set before you pick the model

Every AI project that went sideways on us had the same root cause: nobody could say what 'good' meant until after the demo.

Published: 12 June 2026
min read: 8 min read
Written by: The Attatva engineering team

All insights

The pattern

We have now run enough of these to see the shape. The projects that stalled did not stall on model quality, infrastructure or budget. They stalled because six weeks in, two people in the room had different definitions of a correct answer, and neither had written theirs down.

Why the demo makes it worse

A demo is a curated sample. It is chosen, consciously or not, from the cases the system handles well. Once a demo has been shown to a stakeholder, the perceived baseline is the demo — and every subsequent measurement reads as a regression from something that was never real. Writing the evaluation set first means the baseline is a number that was agreed before anyone had a stake in it.

What the set has to contain

Not a hundred happy paths. Roughly a third ordinary cases, a third genuinely ambiguous ones where reasonable experts disagree, and a third the failure modes you already know about. The ambiguous third is the valuable part: it forces the domain experts to state a policy, and half the time that argument is the actual deliverable of the scoping phase.

Then wire it into CI

An evaluation set that lives in a notebook decays within a month. Ours run on every pull request against a pinned dataset, and a regression fails the build. This is unglamorous and it is the single practice that most reliably separates the systems still running a year later from the ones quietly switched off.

The objection

It costs one to two weeks up front, and it is the first thing cut when a timeline tightens. We now refuse to skip it, and we price it into the scoping engagement so it cannot be traded away later. Every time we have been talked out of it, we bought that fortnight back with interest during hardening.

Tell us what's breaking.

Send the problem, not a spec. We'll come back within one business day with an honest read on whether we're the right team — and who to talk to if we're not.

Email us the problem See how we work

attatvasystems@gmail.com · Replies within 1 business day, Kathmandu time (UTC+05:45)