Worked on the vllm-project/semantic-router repository to deliver a targeted KV cache optimization for Qwen3Guard, focusing on backend development and performance optimization using Rust. The approach involved wiring kv_cache_snapshot() and kv_cache_restore() into the fast path, which eliminated redundant per-request re-prefill of fixed prefixes and improved cache state management. This change reduced end-to-end latency per request and stabilized throughput, aligning with business goals for faster, more efficient inference in semantic routing. The work also included cross-module integration to reuse KV cache APIs, demonstrating a methodical approach to maintainability and resource management in machine learning inference pipelines.