Designing LLM Pipelines for Clinical Data: A Pattern for ALCOA++ and 21 CFR Part 11 Compliance

Most teams shipping LLM features into clinical-data workflows discover the same problem on the same timeline. The first prototype is fast and convincing - a model reads a messy clinical note and produces a clean structured output. Then the questions start arriving. Can you reproduce the run from last Tuesday? Where’s the audit trail? Why did the same input give a different output? What’s the cost at one million records a day? What happens when the model is wrong, and who is accountable? The prototype that answered the first question well rarely survives the rest.