Worked on stabilizing and enhancing the LlamaSwiftKVModel integration within the JetBrains/ArcticInference repository, focusing on robust operation across varying configurations. Addressed a critical bug by implementing conditional logic that checks for FlashInfer availability, ensuring that FlashInfer-specific code executes only when the dependency is present. This approach prevented runtime errors and allowed SwiftKV to function seamlessly even in environments lacking FlashInfer. Leveraged backend development and model optimization skills, primarily using Python, to improve the resilience and testability of the system. The work reduced configuration-related failures and enabled more reliable continuous integration and testing across different deployment scenarios.