In this video
What you'll learn
Diagnose where AI automation breaks
Use five failure patterns from my experiment: distribution, ownership, browser work, memory, and verification.
Design a system with replaceable workers
See how separating control, execution, and verification made each worker replaceable.
Apply five reliability rules to your work
Use factor-based routing, terminal states, destination proof, fresh state, and single ownership.
Why this topic matters
For four months, I ran a digital-products company with almost no routine human execution. AI agents built, marketed, operated, and measured the business. More than 20 people bought a real product. Then I counted my time: I had built myself a minimum-wage job. I’ll walk you through what failed, the architecture that emerged, and five reliability rules you can use in your own AI work.
You'll learn from
Paras Doshi
Head of Data (Amazon, Opendoor)
Paras Doshi led data during Alexa’s early years and through Opendoor’s turnaround after a $1B loss -- two businesses with AI at the center.
Go deeper with a course
Advanced Agentic Analytics in Production & Private 1-1 with Head of Data
Paras Doshi
Head of Data (Amazon, Opendoor)