Skip to content
Work

AI · Evaluation

AI evaluation

Measuring what AI models actually get right: test sets, rubrics, graders and model comparisons.

Overview

Recent work evaluating AI models: building the test sets and rubrics that define good answers, grading model output against them, and turning the results into clear comparisons between models and versions.

What I did

  • Wrote test sets and rubrics for reasoning, coding and instruction-following tasks.
  • Graded and compared model outputs, and flagged failure patterns for training.
  • Built small Python tooling to run evaluations and summarise the results.