The LLM Summary Had to Wait for a Transcript

People talk about summarising a visit as if the model hears the room. In this app it summarised a transcript. No transcript, no honest summary.

That is why sequence mattered

Five APIs and a mismatch was speaker labels, transcript text, and summary drifting. The LLM was downstream.

Retries

If transcribe failed and we still called summary, we invented a visit. React did not. Flutter could not either.

What I do not claim

I am not going to name the model as a brand story. The work was integration: transcript in, summary out, save. The pipeline is the product.