spanish-academy — week of 2026-07-20

A short week, all of it on the conversational assistant.

The biggest fix was language handling. The assistant had been deciding what language to answer in largely by inference inside the prompt, which drifted when a customer switched mid-thread. I moved detection into code and run it on every turn, so the reply follows the customer’s most recent message rather than the tone the conversation opened with.

Related: conversations re-opened after a long gap were dragging old history back in, so someone asking about one course could get answers coloured by a different course they’d asked about days earlier. A conversation now forgets its prior history once it’s been idle past a day.

I also spent time reading real transcripts and turning the gaps into a knowledge update, and added a background pass that re-reads conversations with a stronger model and flags the ones where the assistant answered badly. That gives me a queue to work from instead of relying on spot checks.