I developed the Lumen Inference Engine, a high-performance, low-latency C++ runtime built from scratch to minimize software synchronization overhead in high-frequency AI workloads. Operating on an 8-core CPU architecture, the engine achieves a peak throughput of 295.48 requests per second and an elite optimal tail latency (P99) of 21.28 ms for SqueezeNet models. To completely bypass the traditional kernel-level thread parking and context-switching penalties found in standard mutex-based systems, I implemented a lock-free, bounded Multi-Producer Multi-Consumer (MPMC) queue based on Dmitry Vyukov's algorithm alongside a thread-per-core isolated execution layout. To eliminate global heap contention and memory fragmentation under heavy concurrency saturation, the engine integrates a custom region-based allocator called Lumen Arena. This memory layer facilitates deterministic, thread-local allocation operations using O(1) pointer bumps. Using native Linux kernel performance monitoring units (perf stat), I verified the engine's micro-architectural boundary conditions, demonstrating an optimization jump to 1.265 Instructions Per Cycle (IPC) and a significant reduction in Last-Level Cache (L3) miss rates. This project highlights my core focus on systems programming, atomic data structures, and low-level performance engineering to push hardware capabilities to their absolute physical limits.