Worked on the tenstorrent/tt-inference-server repository to deliver a performance-focused feature by implementing prefix caching in the tt-vllm-plugin, enabling the reuse of previously computed tokens and improving model inference efficiency. Leveraged Python for backend development and model optimization, upgrading vLLM to version 0.18.1 and pinning dependencies such as torch, numpy, and transformers to ensure stability. Addressed code quality by applying lint fixes, removing unused imports, and updating pyproject.toml for better dependency management. The work established a stable production baseline, reducing latency on repeated prompts and enhancing throughput, with a focus on maintainability and repository hygiene throughout the process.