Abstract: Responsive to a request to store a content item for a user of a content service, user metadata and content item metadata are processed with a machine-learned embedding model to obtain a query embedding. A vector similarity search is performed on a vector database with the query embedding to retrieve a stored embedding matching the query embedding, wherein the stored embedding represents a contextual data element related to playback of the content item. A set of inputs is processed with a machine-learned Large Foundation Model to obtain an item storage decision output that identifies a segmentation point at which to segment the content item into first and second segments, wherein the item storage decision output further identifies first and second storage repositories for the first and second segments. The first segment is stored to the first storage repository and the second segment is stored to the second storage repository.

Abstract: A content item is stored to a first storage repository for a user of a content service. The content item comprises content metadata descriptive of content attributes of the content item. An item retrieval probability for the content item is determined based on the content attributes of the content item and historical user retrieval information. A migration threshold value is computed for the content item based on the item retrieval probability, a first storage cost associated with storage of the content item at the first storage repository, and a second storage cost associated with storage of the content item at a second storage repository. The migration threshold value is indicative of a future time T1 at which to migrate the content item. Responsive to occurrence of the future time T1, the content item is migrated from the first storage repository to the second storage repository.

Abstract: Responsive to a request to store a content item for a user of a content service, user metadata and content item metadata are processed with a machine-learned embedding model to obtain a query embedding. A vector similarity search is performed on a vector database with the query embedding to retrieve a stored embedding matching the query embedding, wherein the stored embedding represents a contextual data element related to playback of the content item. A set of inputs is processed with a machine-learned Large Foundation Model to obtain an item storage decision output that identifies a segmentation point at which to segment the content item into first and second segments, wherein the item storage decision output further identifies first and second storage repositories for the first and second segments. The first segment is stored to the first storage repository and the second segment is stored to the second storage repository.

Abstract: Systems and methods for detecting and/or mitigating video artifacts in a video. A processor in a computing device may be configured to receive the video from a video source (e.g., codec or from any other video source compressed or uncompressed) subtract pixel intensities from two consecutive frames in the video to generate a residual, analyze the residual to detect small rectangular shapes, and identifying a video artifact in the video based on a result of the analysis.

Abstract: Various embodiments include systems and methods for performing mid-stream content delivery network (CDN) steering. A processing system in a manifest manipulator may receive a content request from a client device that is directed to obtain content from a uniform resource locator (URL) associated with a primary on-premise CDN. The processing system may also receive an event notification from a CDN decision server that indicates that there is an issue with the primary on-premise CDN. The processing system may modify a manifest file associated with the received content request in response to receiving the event notification from the CDN decision server, and send the modified manifest file to the client device to cause the client device to obtain the requested content from a third-party CDN.

Abstract: Systems and methods for detecting and/or mitigating video artifacts in a video. A processor in a computing device may be configured to receive the video from a video source (e.g., codec or from any other video source compressed or uncompressed) subtract pixel intensities from two consecutive frames in the video to generate a residual, analyze the residual to detect small rectangular shapes, and identifying a video artifact in the video based on a result of the analysis.