Netflix's Two-Tier Workflow Framework: RAW Hollow for State, Conductor for Execution

🎯 Your pipeline treats human review as an exception. Netflix made it a worker type - same slot rules, ML models and AI agents plug into. ⚙️ Pratyusha Singaraju (Netflix) shows the two-tier architecture that makes it work: state and execution kept apart. What's inside: 🔸 Why Conductor alone couldn't answer "how many tasks are waiting on a human right now?" 🔸 Human corrections stored as structured output, not notes - straight into retraining. 🔸 A sub-workflow and a config get you dashboards, QC and human review for free. 🎙️ Pratyusha Singaraju is a Senior Software Engineer at Netflix, where she architects and builds the backend services behind Netflix's content understanding infrastructure - the systems powering search, collections and ratings for millions of viewers. Before Netflix, she worked on Microsoft's data team for Bing, building knowledge graph initiatives across public domains including food, nutrition, politics and events, as well as an enterprise knowledge graph. Connect with Pratyusha: 👉 https://linkedin.com/in/pratyusha-singaraju 📝 Talk: Orchestrating Content Workflows at Netflix Scale | Pratyusha Singaraju This talk was recorded at JSNation Amsterdam 2026: 🔗 https://jsnation.com/?utm_source=youtube&utm_medium=PratyushaSingaraju 📌 Follow the link to watch the full version of all the conference talks, QnA’s with speakers and hands-on workshop recordings: 🔗 https://gitnation.com/events/jsnation-2026?utm_source=youtube&utm_medium=PratyushaSingaraju 🕐 Timestamps: 00:00 - Intro: three kinds of decision-maker, one pipeline 01:12 - Your 20% Off 🫣 01:34 - Rules break, models guess, humans don't scale 04:02 - Two custom workflows, 7 states and 9, nothing reusable 06:21 - The split that made everything else possible 07:35 - RAW Hollow: the entire dataset in every reader's heap 08:45 - Conductor: plug in your sub-workflow, inherit the rest 10:43 - Conductor is powerful. So why wasn't it enough? 11:26 - Onboarding a workflow with config and no code 13:16 - Human in the loop — and why the correction isn't a note 15:07 - The whole flow, end to end 16:29 - The AI agent that opened PRs on the framework running it 17:59 - Four principles that hold it together 📌 We’ve hidden a secret code for 20% off your next event. Join our upcoming conferences! 💫 JSNation US 2026 Nov 16 & 19, 2026 (NewYork & Online) 🔗 https://jsnation.us/?utm_source=youtube&utm_medium=PratyushaSingaraju 💫 React Summit US 2026 Nov 17 & 20, 2026 (NewYork & Online) 🔗 https://reactsummit.us/?utm_source=youtube&utm_medium=PratyushaSingaraju 💫 AI Coding Summit Nov 16 & 17, 2026 (New York & Online) https://aicodingsummit.com/nyc/?utm_source=youtube&utm_medium=PratyushaSingaraju and more! 🚀 Check out all the upcoming events from GitNation: 🔗 https://gitnation.com/events?utm_source=youtube&utm_medium=PratyushaSingaraju Don't forget to use INSIDER20 promo code for 20% off on tickets. #JSNation #GitNation #Netflix #NetflixEngineering #SoftwareArchitecture #SystemDesign #Backend #BackendEngineering #DistributedSystems #Orchestration #WorkflowOrchestration #ConductorOSS #NetflixConductor #RAWHollow #Hollow #StateManagement #EventDriven #EventDrivenArchitecture #Microservices #PlatformEngineering #DeveloperPlatform #InternalTools #HumanInTheLoop #HITL #MLOps #MachineLearning #MLInfrastructure #ModelRetraining #FeedbackLoop #AIAgents #AgenticAI #AIEngineering #ContentMetadata #Metadata #DataEngineering #ScalableSystems