Build an evaluation framework for an AI assistant

Design an end-to-end evaluation framework to catch hallucinations, monitor drift and keep a conversational AI assistant reliable.

Context

The assistant answers customer questions from company documents. Model and prompt changes ship weekly. Wrong answers have caused refunds and complaints.

Clarifying questions to ask first

What does a wrong answer cost?
How often do models and prompts change?
Do we have labelled examples?

How to approach it

Start from the user problem, not the model: is AI the right tool here?
Describe the experience and where the model fits in it
Define what "good output" means and how you would evaluate it before and after launch
Name the failure modes (wrong answers, bias, privacy, cost) and how you contain them
Close with a launch plan and success metrics

Framework: User and problem -> Is AI the right tool? -> Experience design and human-in-the-loop -> Quality bar and evaluation -> Risks, cost and latency -> Staged launch and metrics

Sample answer

Start from failure modes: hallucination, outdated info, refusing valid questions
Layer deterministic checks, LLM-as-judge and human sampling
Every release passes a golden dataset before shipping

Golden dataset from real conversations, labelled by experts
Online: grounded-answer rate, escalation rate, thumbs down, drift alerts
Release gate: no regression on golden set; weekly human review sample

A golden dataset as a release gate, layered checks (rules, LLM judge, human sampling), and live monitoring for drift. Track grounded-answer rate with complaint and escalation guardrails.

What interviewers look for

Defining good output
Offline vs online evaluation
Deterministic checks vs LLM-as-judge vs human review
Release gates and monitoring