Developed a performance and memory-optimized in-memory embedding index workflow for the LEANN repository, focusing on efficient data processing and machine learning operations using Python and NumPy. Introduced a direct index construction method from NumPy arrays, eliminating the need for pickle-based serialization and reducing I/O overhead in ML pipelines. Refactored existing logic to delegate to this new approach, simplifying code paths and improving maintainability. Expanded automated test coverage with seven new tests to ensure reliability across various scenarios. This work enables faster indexing and deployment of embeddings sourced from MLX, GPU, or databases, supporting robust and scalable machine learning workflows.