Shipping an AI Product That Doesn't Lie: A Non-Engineer's Field Report

About this session

I'm a product person, not quite an engineer, and I'm building Lectern: an AI agent that represents speakers and writes venue-specific positioning briefs. This talk is a field report from inside that build, which is still happening today. I'll show how I ship real features as a solo founder working with AI coding agents, and the discipline that keeps a generative product trustworthy. Every factual claim in a brief has to trace back to a confirmed source field. Unsourced claims get surfaced for human review, never silently kept or silently dropped. Prompts are versioned, and an eval suite blocks any change that starts fabricating. You'll see what worked, what broke, and where AI-assisted development still needs a human holding the line.

Speaker

Key takeaways

  • How a non-technical founder ships production features with AI coding agents, and the specific points where that approach breaks down.
  • A concrete pattern for keeping generative AI honest: source-grounded claims, a [VERIFY] surfacing step, versioned prompts, and evals used as a release gate.
  • How to tell AI-generated confidence from AI-generated fact, and where to keep a human in the loop.

Related sessions