TLDR Tech / Dev
How Netflix is Using Multimodal AI to Power Video Search
Netflix lets editors search thousands of hours of raw footage by running multiple specialized AI models: character recognition, scene classification, dialogue, and object detection. A decoupled three-stage pipeline handles it: raw annotations land in Cassandra untransformed, an offline job uses temporal bucketing to merge all model outputs into aligned one-second intervals, and those fused buckets get indexed in Elasticsearch as nested documents.
Read the full story