Worked on the alibaba/MNN repository to enhance LLM engine integration by adding direct input embeddings support and refactoring the forward path for improved maintainability. Leveraged C++ and embedded systems expertise to introduce overloads that allow input embeddings to be passed directly, enabling more flexible generation workflows. The technical approach involved simplifying the processing logic by removing redundant parameters and delegating responsibilities within the forward path, which reduced technical debt and streamlined future enhancements. These changes improved API cleanliness, reduced the risk of regressions, and provided greater control over embedding-based prompts, aligning with performance and reliability objectives for LLM applications.