Libro

Every AI demo assumes the model answers. Every production system assumes it won’t. This book is an engineering account of the difference. Inside: the layered architecture of a platform that routes 24 intents across 23+ engines and decides, per request, whether quality justifies aggregation. A reliability layer with three redundant gateway proxies, preemptive credential rotation at 95% of limits, and a 20-slot failover chain — stress-tested to full credential exhaustion with no user-visible outage. A seventeen-intent security engine that treats disagreement between four models as a signal rather than noise. And a continuous self-training loop that turns production traffic into a better model: 857,000 instruction pairs, LoRA fine-tuning, validation loss down from 1.42 to 1.31. Measured on the running system: 4,117 requests over 30 days, 99.2% fast-tier completion, p95 latency of 7.8 seconds, and a corpus that grew by 115,305 pairs in a month. And the parts that failed: a dataset manifest that pointed at repositories which did not exist; a scraping pipeline humbled by JavaScript single-page applications; and an A/B study the book openly has not yet run.