# Production AI Models Run on a 90-Component End-to-End Infrastructure Ecosystem

This talk explains the production infrastructure required to make AI models work reliably at scale, arguing that the model itself is only one part of a much larger system. Using a social media feed as the example, it traces a single user request from candidate retrieval and real-time feature lookup through ranking models, policy and safety checks, and final response assembly, all within a tight latency budget where caches, deadlines, fallbacks, and partial results may be necessary. It then follows the feedback loop after serving, covering event logging, streaming platforms such as Kafka, schema validation, data lakes, and the creation of online and offline features with parity checks to prevent drift. The talk also outlines training and release operations, including versioned datasets, orchestration, checkpoints, experiment tracking, evaluation, model registry, and gradual rollout through shadow traffic and canaries. The central lesson is that successful AI products depend on coordinated serving, data, training, infrastructure, observability, and reliability systems.