Insights and Epic Fails from 5 Years of Building ML Platforms

Learn from five years of real-world experience building ML platforms that serve over 14 million YouTubers and Amazon's largest third-party seller in this candid conference talk. Discover the architectures, strategies, and critical failures that shaped successful MLOps implementations across three different platform builds. Explore practical insights on tool selection using the "9 jobs to be done" framework, understand why drift monitoring is often overrated while data quality issues pose the real threats, and learn when offline inference outperforms endpoint serving. Gain valuable perspectives on balancing data science autonomy with engineering rigor, implementing effective data lineage to prevent target leakage, and why medium-sized data tools frequently outperform over-engineered technology stacks. Examine real-world tradeoffs between cloud GPU services and on-premises alternatives, and understand how to design ML platforms that achieve genuine adoption, stability, and measurable business value rather than falling into common MLOps hype traps.