I talk about the same thing I write about: delegating more to AI while staying in control, by building the cheapest check that tells you whether it worked.
Want me to speak at your event? Reach me on LinkedIn.
Before you let an AI agent work unsupervised in a domain, give it an exam: versioned questions with assertions that say what good enough means.
On the cost and time savings of using AI personas alongside surveys. The original title isn’t public.
Getting an agent to do something is easy now. Making it cheap to change takes executable checks, the right box, and sometimes no inference at all.
What economists delegate to AI coding agents, what they never delegate, and how exam-first iteration keeps them in control.
Why the Airflow community needs an agreed-upon way to benchmark LLM tools and agents, and what that could look like.