Worked on enhancing the langchain4j repository by addressing memory management and resource lifecycle issues in ONNX-based scoring models. Focused on Java, the developer implemented AutoCloseable interfaces for core components, enabling safe and explicit cleanup of native resources such as OrtSession and HuggingFaceTokenizer. By introducing close() methods and leveraging try-with-resources during initialization, they prevented memory leaks and reduced the memory footprint in long-running production workloads. The work included extending unit and integration tests to ensure robust regression coverage, ultimately resolving a memory leak during startup and delivering a more stable and maintainable scoring pipeline without breaking existing APIs.