feat: I built a cyberpunk cat RPG to teach production AI architecture

Show Notes

I built a cyberpunk cat RPG where you play a TabbyWarrior or a SphinxRogue fighting through Neo-Pawsburg in 2087, and right next to the game, you can watch the real production AI architecture executing live underneath it. I built this for AWS Summit LA because I got tired of watching people's eyes glaze over at architecture diagrams.

This episode walks through the five reliability patterns that actually decide whether an AI agent survives production: grounding the model with retrieval before it speaks, retries and idempotency without breaking randomness, a hard boundary where the model proposes and the system disposes, structured observability from the first Lambda invocation, and state that lives in a database instead of a prompt.

Then I tell you about the seven hundred dollar mistake I almost made. "Serverless" does not mean free at idle, and I found that out the expensive way with OpenSearch Serverless's pricing floor, one day before AWS fixed the exact problem I'd already engineered around.

If you've ever needed a straight answer on what "the AI just calls an API" is actually hiding, or you build agentic AI systems and want the reliability patterns without the marketing, this one's for you. None of this is because AI is uniquely broken. It's because an AI agent is a distributed system, and distributed systems have rules.