Owners of busy restaurants rarely reply to their Google reviews. It is forty minutes of work at eleven at night, and it is the first task to be dropped. Titanium drafts the replies for them in the voice of that particular business, and the owner reads them and sends.
Getting a language model to write one convincing reply takes an afternoon. The harder problem is the hundredth reply, which still has to read as though a person sat down and wrote it.
Four passes
A reply is drafted, and a second pass then critiques that draft and rewrites it. We found the second pass mattered more than expected. A model told to avoid a phrase will drop the words and keep the pattern, so the check has to read the output itself. After that the reply is trimmed to a word budget that shifts with the star rating, and checked against a short list of rules that cannot be broken. One of them is that it may never invent a fact. It cannot name a dish the reviewer did not mention.
Variation is enforced outside the model
Google removes owner responses that mirror each other across a profile, and a model has no memory of what it wrote yesterday. So every reply that is sent gets recorded. The last thirty are analysed for repeated openings, closings and three-word runs, and those are handed to the next draft as phrases it may not use.
We put that control outside the model because the model cannot be relied on to police itself. It is the reason the system can be left running.
Measured on a fixed test
Every version is scored against the same twenty-five reviews, before and after each change to the prompts. On that set the replies went from 3.5 out of 5 to 4.8.
All work