AI / ML
LLMs in Production: Structuring Unstructured Speech
By Vivek Jalondhara · 2 min read
Published Nov 2025 · Updated
Validate the shape of model output, preserve its source and design a review path for unsupported or missing information.
Structured output makes it easier for an application to consume a model response. A schema can describe permitted fields and missing values, but it cannot prove that an extracted statement was present in the source. Keep source references available to the person reviewing the result.
Validate every response before storing or displaying it as an accepted record. Reject invalid shapes, handle timeouts explicitly and allow incomplete output to enter a review state. A model-generated confidence score is not a substitute for measured accuracy or human review.
Version prompts and evaluate changes against suitable, authorized examples. The walkthrough here describes integration concerns; my Ecoscript contribution was the React frontend, browser recording and transcript/summary management, rather than training or validating the clinical models.
Keep generated summaries reviewable
For an application that manages generated summaries, distinguish a draft from a reviewed record. This is an illustrative workflow model, not a clinical recommendation or a production claim.
Illustrative example · TypeScript · Simplified for explanation
type Summary = {
transcriptId: string;
text: string;
review: "pending" | "reviewed";
};
const canFinalize = (summary: Summary) => summary.review === "reviewed";- Transcript
- Generated draft
- Validation
- Authorized human review
Conceptual flow for the example above.
Tradeoff to keep in mind
A valid data shape does not guarantee that the content is correct. Keep the source available for review and avoid treating a model's output as a verified fact.