AI product designJune 16, 2026
AI PMs Should Start With the Workflow, Not the Model
A practical case for defining the user workflow, decision, and operating constraint before choosing a model, agent, or automation pattern.
Insights
Public explanations for the ideas behind the work.
These pieces show how I frame AI product problems in public: start with the workflow, understand the consequence of a wrong action, and treat production behavior as part of product design. Each featured item links directly to the original source.
Featured
Two useful starting points
Production systemsMay 7, 2026
Why AI Systems Fail in Production
Moves the failure analysis beyond model quality and into the surrounding system: validation, execution, traceability, and the action that follows an output.
Selected writing
From model output to product behavior
The common thread is practical rather than ideological: what should the product do next, and how should that behavior be measured?
June 4, 2026
Reliable AI Is a Product Design Problem
Connects model behavior to product mechanics such as review, escalation, runtime controls, auditability, and operating metrics.
May 28, 2026
Human Review vs Automation in AI Systems
Argues for risk-based review at specific control points instead of putting a person in every loop or automating everything by default.
May 21, 2026
Validation Before Action
Explains why a plausible model response still needs action-specific validation once the output can change production state or business workflow.
Case study
ReplayGuard is where the thesis becomes a working artifact.
ReplayGuard combines an issued patent, public implementation, browser-local evaluation, technical limitations, and commercialization material in one place. It is the strongest current example of the product argument and the underlying implementation living together.
Open ReplayGuard →