An engineering blog from Diffio detailing how its speech and audio restoration models are deployed and optimized, including Mac mini inference fleets, Neural Engine partitioning and cold-start behavior on Modal. Issues present measured benchmarks, tables and candid negative results.