Jasmine Sarabia

Jasmine Sarabia

Back to Work
AI Quality / Evaluation Systems

AI Model Evaluation System

Designing a repeatable way to test whether AI is accurate, useful, consistent, and worth deploying.

As AI moves from experimentation into real workflows, quality can’t be judged by whether an output simply looks good in a demo.

The question becomes: what evidence is enough to trust the system?

A practical way to evaluate AI

I organized evaluation around four decisions teams repeatedly need to make.

01Correct

Is it correct?

Accuracy · grounding · source alignment

Can we trust the answer?

02Useful

Is it useful?

Human evaluation · acceptance · edit rate · trust

Does this actually help someone do the work?

03Consistent

Is it consistent?

Regression testing · eval gates · drift detection

Will quality hold as the system changes?

04Worth It

Is it worth it?

Cost · latency · complexity · model comparison

Is the additional quality worth the operational cost?

From model change to ship decision

Evaluation works best when it becomes part of the workflow rather than something teams do only when something goes wrong.

01

Define

Define the use case

What is the system supposed to do, for whom, and what would failure mean?

02

Test

Build the evaluation set

Create representative prompts, tasks, questions, and expected outcomes.

03

Score

Define the rubric

Accuracy · grounding · usefulness · tone · safety · trust

04

Compare

Run models against the same standard

Compare models, prompts, sources, or workflow versions using the same evaluation set.

05

Calibrate

Check automated scoring with people

Make sure the evaluator is rewarding the right things.

06

Decide

Translate performance into a decision

SHIPHOLDREVISETEST FURTHER
07

Monitor

Keep testing after launch

Watch for quality changes, new behavior, and drift over time.

The evaluation toolkit

Before launch

Quality & grounding

After launch

Model decisions

Before launch. After launch.

Eval gates guard the door. Drift detection watches the room.

Before launch

Eval Gate

Triggered by a change:

  • new model
  • new prompt
  • new source material
  • new workflow

Does the new version clear the bar?

SHIPHOLDREVISE

After launch

Drift Detection

Runs even when the team changes nothing. Looks for:

  • declining quality
  • changing user behavior
  • model changes
  • trust erosion

Is performance changing over time?

MONITORINVESTIGATERECALIBRATE

Quality is bigger than accuracy.

A system can be technically correct and still be difficult to trust, expensive to operate, or unhelpful in practice.

Conceptual example — not production data

92%Accuracy
88%Grounding
61%Trust
$4.80Cost / 1K outputs
HOLDDecision

Accuracy improved, but grounding still needs review and users continue editing a meaningful share of the output.

Decision → revise before broad launch.

A metric is only useful if someone knows what to do with it.

Grounding +4%

The AI produced source-supported answers in four more cases out of every hundred.

Judge model shows length bias

The scoring system may be rewarding longer answers even when they aren’t better.

Drift check failed

The AI is performing worse than it did previously even though the team didn’t intentionally change the workflow.

What this system enables

  • Repeatable quality standards
  • Model comparison
  • Human-calibrated evaluation
  • Clear ship / hold decisions
  • Post-launch monitoring
  • Governance-ready evidence
  • Plain-language reporting

The goal isn’t perfect AI.

It’s knowing what good means, measuring it consistently, and knowing what decision to make next.