frunt turns a restaurant's own documentsinto answers and training its staff actually use.

What happened

A restaurant's knowledge lives in PDFs, photos of laminated sheets, spreadsheets and paper. None of it is written for searching.

What I chose

Claude reads every document, photos included, so there's one engine instead of an OCR tool plus a model (Tesseract, Textract and Mistral OCR were the alternatives). Documents are split into passages and stored with their meaning in Postgres (pgvector), every restaurant in one database, kept apart by row-level security; a database schema per restaurant would mean every change runs once per customer.

How it turned out

Upload, ask, cited answer worked end to end on 18 May, three days after the first commit. Staff could ask and train on their phones four weeks after it.

"I'm proudest that the base product works. Not a technical feature: the goal I set out to achieve."

What happened

Allergen questions were answered like any other: search the documents, let the model phrase the answer. A confident "the brownie is nut-free" when the matrix says otherwise is a safety incident.

The options

A second AI to check the first. Refuse every allergen question. Or a plain-code check.

What I chose

Allergen facts come from the database, read straight from the allergen matrix the manager confirmed, never from the model. A plain-code check reads every answer that names an allergen: it either adds the verified facts or replaces the whole answer with a refusal that points to the matrix.

Why not a second AI

"A checker sharing the generator's model shares its blind spots." The cases the first model gets wrong are the ones a second model gets wrong too.

How it turned out

The model phrases the answer; it is never the source of an allergen fact.

What happened

I came back from six weeks away and found three things that had broken quietly. A one-letter typo in a setting name meant payment updates had never reached the database since launch. Scheduled jobs added after 1 June had never run once, because the job service had stopped syncing. Nothing had errored. Nothing had said anything.

What I chose

Every scheduled job now records a run every time, even when there's nothing to do, and a weekly email says green or red.

The rule

"Nothing to do" has to look different from "did not run".

What happened

frunt drifted. I'd built outward from the answers into a rota and team briefs, and by summer the manager dashboard opened on "Build next week's rota", with the answers third in the menu. Every step made sense on its own.

The options

Delete the rota and briefs. Rename the product to something that explains itself. Turn it into an open-ended assistant.

What I chose

Freeze the rota and briefs (switched off, not deleted) and put the answers first. And a third test that every new feature must pass: do the restaurant's own documents make a hard job easy? The rota fails it: it never touches a document.

How it turned out

It shipped the same day: the rota and briefs switched off by default, and the website leading with the answers. The manager home became one screen on 28 Aug.

"Selling the product became confusing because it was third in the list of features when you open the app."

What happened

Three things found by testing frunt the way a venue would:

• Over-citing. "What are the Christmas hours?" cited seven documents, none of which answered it. Answers now cite only the documents they used. (Live, 28 Sep.)
• Lost pages. A 40-page handbook silently lost pages 19 to 40, because text extraction for long PDFs had never been built. Now it has. (Live, 23 Sep.)
• A planted mistake. In a dry run on the live app, I planted a wrong cooking temperature in a test document. frunt quoted it back, with a citation. A citation proves where an answer came from, not that the document is right. So frunt's document checks are an aid, and the manager stays responsible for what their documents say. (Written as ADR 0084 on 24 Sep.)

What happened

I built a fixed test for frunt's answers: 38 questions asked of a restaurant I invented, marked by code, not by a model. Its first run failed three of the eight safety questions. Asked about a dish missing from the allergen matrix, frunt marked the model's guesses as verified instead of refusing.

What I chose

Fix the rule, not the wording. The list of every dish with a given allergen now only answers questions that ask for a list. A question about one dish the matrix doesn't have gets no facts, so frunt refuses it. Fixed and re-tested the same day.

The test now

31 of 32 scored answers right, safety floor 8 of 8, on a fixed set of 38 questions marked by code (7 October 2026).