Exa, a modern AI search engine, has enhanced its real-time web content extraction and embedding capabilities through significant engineering efforts. Initially challenged by Python's Global Interpreter Lock (GIL) limiting CPU utilization, Exa transitioned to Rust, allowing for better parallelism and quadrupled throughput by efficiently dividing tasks between CPUs and GPUs. The transition revealed bottlenecks in synchronous operations and inter-process communication, which were alleviated by Rust's parallel iterators. However, this led to new challenges such as CUDA out-of-memory errors due to thread safety issues, which were resolved by refactoring data handling. The improvements exemplify Exa's commitment to building scalable, performance-critical systems, inviting others to join their innovative projects.