WorkshopSeptember 26, 2026

Building an AI system you can actually measure

A minimum viable evaluation setup for RAG and assistants, and what to do when the numbers say your design decision was wrong.

What we cover

I benchmarked my own retrieval default and it lost. We build the same kit: a hand-labeled question set, an ablation across configurations, answer-level evals with the model on and off, and latency and cost from traces.

Conferences, meetups and internal engineering days are welcome. Invite me through the contact page.

Built from

The Alex OS evaluation report.