A deep technical analysis of a self-generating podcast pipeline that reveals systemic reliability issues hiding in plain sight. This episode examines scattered retry logic across multiple workers, the dangers of hand-rolled backoff strategies, and the architectural debt accumulating as AI model calls compound latency. Hosts dissect critical decisions including the Haiku/Opus model split for cost optimization, idempotency fixes in Durable Objects, and the state machine complexity of podcast lifecycle management. Key insights include the risks of hardcoded model names without centralized configuration, the need for consistent idempotent endpoints across distributed services, and how feature velocity masks infrastructure debt. Perfect for engineers building production AI systems who need to understand when tactical fixes signal deeper architectural problems. Topics covered: exponential backoff patterns, LLM pipeline latency optimization, Durable Object state machines, observability instrumentation, and the transition from prototype to product infrastructure.