Transcribing a few audio files with Whisper is easy; doing it for millions of recordings reliably and cost-effectively is another problem entirely. This talk digs into the Python code and infrastructure behind a large-scale speech-transcription platform built for the insurance industry, tracing how a notebook prototype grew into a distributed inference pipeline running across thousands of GPU workers. The focus is inference engineering rather than ML theory: benchmarking CPU vs. GPU workloads, maximizing throughput, orchestrating jobs with Azure ML, surviving spot-instance interruptions, and writing resilient code that recovers from failures and resumes automatically. Expect real benchmark results, architecture decisions, code examples, and lessons learned from production.