CrackML by @ml.with.umang
GenAI interview guides / LLM Evaluation Interview Guide
GenAI / LLM

LLM Evaluation Interview Guide

Build evaluation systems for generative AI that combine task metrics, model graders, humans, and online signals.

Failure taxonomy

Define product-specific failures before choosing a score. Averages can hide hallucination, safety, latency, or task-completion regressions.

Offline evaluation

Use representative datasets, adversarial slices, deterministic checks where possible, and rubric graders calibrated against humans.

Online evaluation

Track user outcomes, corrections, abandonment, latency, cost, and sampled review. Separate model changes from traffic-mix changes.

Release discipline

Version prompts, models, tools, data, graders, and policies; keep important regressions as permanent tests.