A Netflix data scientist recently spent three months building a viewer embedding model from scratch—only to discover, by accident, that the Studio team had already created a far more sophisticated content embedding for analyzing scene structure. The overlap was invisible because the two teams operated on separate tech stacks, separate registries, separate interfaces. Neither knew what the other had built. This is not a story about one team's inefficiency. It is a story about what happens when machine learning scales across an entire company without the infrastructure to connect it.
Three years ago, Netflix applied ML primarily to personalization—recommendations, engagement, member experience. Today ML runs across personalization, studio workflows, payments, advertising, and a growing list of business domains, each with its own tech stack, metrics, and organizational structure. The diversity is a success. The fragmentation is the cost.
"Without any discovery infrastructure, ML practitioners couldn't easily collaborate or share work across business verticals," wrote Netflix engineers Saish Sali, Nipun Kumar, and Sura Elamurugu in a post published Monday. The company's own tools existed in silos. The model registry didn't know which A/B tests were running on its models. The pipeline orchestrator couldn't see downstream dependencies. Finding a model meant opening one system; understanding its lineage meant switching to another; tracking experiments required a third. Practitioners couldn't answer basic questions: What features already exist? What data sources are available? Has someone solved this problem before?
Netflix's answer is the Model Lifecycle Graph, a system that links three previously disconnected platforms—the model registry, the A/B test experimentation platform, and the pipeline orchestrator—into a single graph where relationships between models, experiments, and data pipelines become visible and traversable. A practitioner can now trace where a model came from, which experiments use it, and what downstream systems depend on it, without context-switching across five different interfaces. The Studio team's content embeddings—originally built to identify scene boundaries and visual transitions—suddenly become discoverable by the Ads team, which needs context matching to align advertisements with what viewers are watching. The same embeddings could improve episodic merchandising in personalization, matching episode mood to user preference.
Netflix is now running the Model Lifecycle Graph across 1,000+ internal models. The company frames this as ML democratization: turning machine learning from a specialist tool used in isolated pockets into company-wide infrastructure any practitioner can access. The economic logic is straightforward. Duplicate model development wastes compute, engineering time, and opportunity. When models are invisible, teams reinvent them. When they are discoverable, teams reuse them.
But the broader implication extends beyond Netflix's walls. The company predicts that within three years, every large organization will face the same problem it faced: AI assets accumulating faster than the organizational structures to connect them. The companies that solve this first will have a structural advantage—not because they have better models, but because they can use the models they already have. The Model Lifecycle Graph is early infrastructure for a world where AI capability is no longer the constraint. Organization and discoverability are.