Worked on enhancing text embedding pooling strategies in the opensearch-project/ml-commons repository, focusing on support for decoder-only models and pre-pooled embeddings. Developed the LAST_TOKEN pooling method to extract the last non-padding token embedding, improving context handling for specific model architectures. Introduced a NONE pooling mode to efficiently bypass pooling when sentence embeddings are pre-computed, increasing flexibility and reducing latency. Expanded validation logic and implemented comprehensive unit tests to ensure correctness across ONNX and TorchScript formats. Updated documentation and release notes to communicate these changes, leveraging Java, Markdown, and machine learning expertise to deliver robust, production-ready features with clear business value.