Building AI That Ideates from the Outside-In

About this session

Most "AI for creativity" tools are wrappers, and have the same failure mode: ask an LLM for a fresh idea and it hands you the most statistically average one. The prompt is asking for creativity as if it were a vibe, not a mechanism.

Real innovation must be transferrable. Someone took a solved problem from one domain and smuggled it into another where nobody had made the connection yet. That's a structural move that can be engineered.

This talk walks through Pattern Thief, an AI system built specifically to do that: force genuine cross-domain analogical retrieval instead of same-domain remixing. It'll cover the actual build: a 28-domain taxonomy, a retrieval layer that surfaces precedents from unrelated fields, and structured "Go Deeper" prompts that turn an abstract analogy into something implementable.

Plus the specific design decisions that separate "make something creative" prompting (which fails) from constraint-based cross-domain retrieval (which doesn't).

Expect a live demo, the prompting and retrieval architecture behind it, the dead ends that didn't work, and a broader argument for builders: if you're shipping an AI product meant to expand what people can think of, the interesting engineering problem isn't the model, it's the scaffolding.

Speaker

Key takeaways

  • A concrete architecture pattern — domain taxonomy, precedent retrieval, structured output — for building AI tools that expand what users can think of, not just automate what they already do.
  • See exactly why "give me something creative" prompting fails, and the specific constraint-based retrieval approach that replaces it — reusable in your own product the same week.
  • Leave with a sharper lens for evaluating your own AI roadmap: is it optimizing within a domain, or actually crossing one? Most AI product teams are shipping faster averages. The difference between that and shipping something nobody's seen before.

Related sessions