Auguste

A chef’s library that answers back

Auguste

211 cookbooks and a chef’s own recipes, in one place you can ask. Every answer shows the book and the page.

Culinary · Minor Hotels Thailand · October 2026
Scroll or press →
Auguste

The problem

The knowledge is on the shelf. Finding it is the hard part.

1

A serious cookbook shelf holds more than any chef can remember.

2

Search by title finds a book. It does not find an idea.

3

The best ideas sit between books: a technique from one cuisine, a pairing from another.

Auguste turns the shelf into something you can ask.
Auguste

What it is

Ask in plain words. Get an answer with its sources.

How do I stop a whole fish sticking to the grill?

Ah, the oldest enemy of the grill master — the fish that leaves half its skin on the bars.

The Vietnamese way, on a cast-iron griddle set on the coals 1 … Lay the fish down and do not move it for about 6 minutes, until the skin is deep golden.

The direct-grill way, on an oiled hinged rack 2 … Broil turning once, exactly 10 minutes per 2.5 cm of thickness at the thickest part.

The one critical point: dry skin plus a hot, oiled surface. Wet skin steams, sticks, and tears. Dry it, then dry it again.

1 Vietnamese Home Cooking (Charles Phan) · p. 3512 James Beard’s American Cookery · p. 147
A real answer from Auguste, shortened (…). The cards it used are named under it.

  • Search

    Recipes, menus and books, by name, ingredient, technique or course.

  • Ask

    A chat that answers from the library and names the page.

  • Pair

    A map of ingredients: what sits close is used together.

  • Check

    Allergens are estimated on every card. Always check before use.

Auguste

What is inside today

A library you can ask.

211
cookbooks, read into the library
84,526
pages read
60,379
recipe cards, one dish on each
4,501
ingredients on the pairing map
602
menus, the chef’s own and others’
1,095
of the chef’s own recipes, filed like the rest
All of it can be searched by name, by ingredient, by technique, or by meaning.
Auguste

How it was built

Eight steps from shelf to answer.

NO. 01
Read
Every book becomes text. Poor scans are re-read from page photos.
25.3 million words
NO. 02
Cut
The text is cut into recipe cards: one dish, one card. Quantities are converted by code.
60,379 cards
NO. 03
Label
An AI gives each card a dish name, course, cuisine, technique and allergen estimate.
14 allergens
NO. 04
Check
A second AI re-reads a sample. Cards that fail are fixed or rebuilt.
834 of 2,721 failed
4,380 rebuilt
NO. 05
Map
Recipes are placed by meaning. Ingredients are linked when books use them together.
93,725 pairs found
NO. 06
Answer
A question fetches the best cards. The AI writes the answer from those cards only.
about $0.003 a question
NO. 07
Voice
A character, a spoken voice and an ear. Talk to Auguste while you cook.
2.6 seconds to a first sentence
NO. 08
Keep
Nightly backups and a private cloud copy, so nothing is lost.
119,247 files protected

AI did the labour. The chef decided what good looks like, and checked the results.

Auguste

The rule that makes it trustworthy

The library knows. The AI only reads and explains.

Question

Asked in plain words, the way you would ask a colleague.

→

Look up

The best cards are found by words and by meaning. Search typically takes under 48 milliseconds.

→

Read

The AI is given only those cards, plus strict rules. It does not answer from memory.

→

Answer

A short answer. Each claim is marked with the card it came from.

If the library has no answer, Auguste says so.
Facts come from the cards.Quantities are the book’s, not the AI’s guess.
Every claim cites a card.You can open the card and the page.
Opinions are marked.“My opinion, clearly marked as such.”
Auguste

The map

61,474 recipes on one map.

Each dot is a recipe. Recipes with a similar meaning sit close together. Colour shows the course.

    Move over the map to light up one course.

    A second map links 4,501 ingredients that books use together.

    Auguste

    The adviser

    A senior adviser, with a character.

    1

    An opening line

    Ah, the oldest enemy of the grill master — the fish that leaves half its skin on the bars.
    2

    A clean answer

    Method and quantities, from named books and pages.
    3

    A closing gift

    A trick from the old days: before service, run a cut onion or a folded cloth dipped in oil over the bars … If it hisses and skates, you are ready.
    He speaks and listens
    Bonsoir, Chef.

    A calm old-Frenchman voice, picked from 17 candidates. Press the phone icon and talk while your hands are busy. The first sentence comes back in about 2.6 seconds.

    In the spirit of Auguste Escoffier, senior adviser to every kitchen.

    Auguste

    How long it took

    20 days from the first book to a live service.

    22 Sep · Day 1

    Books are read into text.

    23 Sep · Day 2

    Recipe cards cut. An AI labels every card.

    24 Sep · Day 3

    Search, chat and site built. Live behind a PIN by mid-afternoon.

    25 Sep · Day 4

    The own-model experiment. Ingredient pairing opens. A voice is chosen.

    26 Sep · Day 5

    The chat engine is chosen. Ingredient pictures are drawn.

    27 Sep · Day 6

    A private cloud copy and nightly backups.

    28 Sep · Day 7

    Dishes tab and saved chats: 37 changes shipped in one day.

    29 Sep · Day 8

    The Mac Studio travels to Thailand.

    6 Oct · Day 15

    Back online in Thailand.

    8 Oct · Day 17

    Reference notes: research can sit beside the books.

    11 Oct · Day 20

    Its own address, auguste.mikarina.org.

    Built before starting a new role, and in spare hours since.
    Auguste

    Who did what

    One chef, an AI team and one rule: the chef approves.

    The chef

    Decides what it is for. Reads the results. Approves or rejects.

    About 1 in 5 ingredient pictures was rejected by eye.

    The builders

    Claude Code writes the software, one small change at a time. Tests must pass before a change goes live.

    241 commits and 50 test files since 25 Sep.

    The workers

    MiMo reads, labels and answers. Gemini read the books. Models on the Mac Studio draw pictures, speak and listen.

    Free models run on the Mac itself.
    $368
    of AI usage for the first version, 22 to 26 Sep. $160 of it is an estimate of coding time.MiMo $106 · Gemini $97 · Claude API $5
    $0.003
    a question since, on the chat engine.
    Auguste

    What went wrong

    Four things that broke, and what they taught.

    Old scans garbled the quantities.

    A quarter cup was read as “%4”.

    Now48 worst books re-read from page photos. Quantities are converted by code, not by the AI.

    A model trained on the chef’s writing lost.

    In 86 blind comparisons it was never chosen. The longer, better-sourced answer won every time.

    NowThe knowledge lives in the library, not in the model.

    Credit ran out in the middle of a job.

    After 171 of 206 books, and again with another AI.

    NowEvery paid run starts with a small pilot and a spending cap.

    “Everything is broken”, while the tests passed.

    A layout saved on a 3,392-pixel screen was applied to a 1,666-pixel window.

    NowTests run in the real browser with the real saved settings.

    Still open: a second training round is paused. It needs more memory than this Mac has.
    Auguste

    For our kitchens

    What it can do for us.

    Works today

    • Menu development. Search 602 menus and 60,379 recipes by ingredient, technique or cuisine.
    • A research desk. Notes for current projects sit beside the books and are cited like a book.
    • Technique questions answered with the page. Useful for training.
    • Pairing ideas for new dishes, from the ingredient map.
    What would you ask it first?
    Auguste

    Private by design

    Private, backed up, and honest about its limits.

    PrivateBehind a PIN. Not published, not indexed, not shared.
    Runs on one Mac StudioIn Thailand, with a private cloud copy as a safety net.
    Not connected to Minor systemsIt has no link to any hotel system.
    Backed up every nightCopies in more than one place.
    Questions leave the MacTo write an answer, the question and the matching cards or notes go to the AI provider.
    Allergens are estimatesMarked as estimated. Check before use. Always.
    Auguste

    Try it

    Ask it something.

    auguste.mikarina.orgAsk Mickael for the PIN.
    Mickael Farina · Cluster Director of Culinary · Minor Hotels Thailand
    Appendix

    Appendix 1

    How one answer is made.

    1 · Question

    You type or say it in plain words. Recent turns of the chat go with it, so follow-up questions work.

    2 · Two searches at once

    One by words (SQLite full-text search), one by meaning (BGE-M3 vectors). The best cards from both are merged. Notes you add to the library are searched too.

    3 · Strict rules

    The model gets the cards and a rulebook: facts only from the cards, cite each as [n], say “not in the library” when it is not, mark opinions, estimate allergens.

    4 · The model writes

    MiMo V2.6 Pro, a cloud model, receives the question and the cards and writes the answer in the adviser’s voice. If it fails, a local model on the Mac answers instead.

    5 · The page shows it

    The answer appears with the cards it used. Open a card to see the page it came from.

    6 · It is saved

    The chat is kept in Past chats, on the Mac.

    Appendix

    Appendix 2

    What it is made with.

    Builders

    • Claude Code writes and tests the software.
    • Claude in Chrome checks the pages in a real browser.
    • GitHub keeps every change; the Mac updates itself from it.

    Paid AI

    • MiMo V2.6 (Xiaomi) tags, audits and answers.
    • Gemini Flash (Google) read the books.
    • Claude API did card batches and audits.

    Free, on the Mac

    • Qwen3.6-35B for tests and training.
    • BGE-M3 for search by meaning.
    • Qwen-Image 2.1 drew the ingredient pictures.
    • Kokoro speaks, Whisper listens.

    Serving

    • FastAPI + SQLite run the site.
    • Cloudflare tunnel, Workers and storage.
    • Mac Studio (Apple M1 Ultra, 64 GB).
    • Backups to several places, every night.
    Appendix

    Appendix 3

    Where every number comes from.

    FigureWhat it isHow we know
    211Cookbooks in the libraryCounted from the library folder, 11 Oct 2026.
    84,526Pages read from the cookbooksCounted with a PDF reader, 11 Oct 2026.
    25.3Words of text read (millions)Counted from the extracted book text, 11 Oct 2026.
    60,379Recipe cards from the booksCounted in the library database, 11 Oct 2026.
    1,095The chef’s own recipesCounted in the library database, 11 Oct 2026.
    602Menus, the chef’s own and others’Counted in the library database, 11 Oct 2026.
    4,501Ingredients on the pairing mapCounted from the pairing data, 11 Oct 2026.
    61,474Recipes on the mapCounted from the map data (cookbook cards plus the chef’s own), 11 Oct 2026.
    11,996Dots drawn on these slidesA sample of the map, evenly spaced within each course.
    14Allergens estimated on each cardThe fixed list the cards use.
    93,725Ingredient pairs found in the booksNoted in the build log, 25 Sep 2026.
    2,721Cards re-read in the quality checkNoted in the build log, 24 Sep 2026.
    834Of those, cards that failedNoted in the build log, 24 Sep 2026.
    4,380Damaged cards rebuiltNoted in the build log, 24 Sep 2026.
    48Worst scanned books re-read from page photosNoted in the build log, 26 Sep 2026.
    119,247Files protected by the nightly backupNoted in the handover, 28 Sep 2026.
    0.003Cost of one question (US dollars)Measured on the chat engine’s spending log, 11 Oct 2026.
    2.6Seconds to the first spoken sentenceTimed in a voice call, 25 Sep 2026.
    FigureWhat it isHow we know
    17Voices compared before one was chosenNoted in the build log, 25 Sep 2026.
    368AI usage for the first version (US dollars)Stated total for 22 to 26 Sep 2026. Part of it is an estimate.
    160Of that: coding sessions (estimate)The chef’s own estimate. Not measured.
    106Of that: MiMo (US dollars)Noted in the build log, 22 to 26 Sep 2026.
    97Of that: Gemini (US dollars)Noted in the build log, 22 to 26 Sep 2026.
    5Of that: Claude API (US dollars)Noted in the build log, 22 to 26 Sep 2026.
    86Blind comparisons, own model against plain modelNoted in the build log, 25 Sep 2026. The own model was never chosen.
    171Books read when the credit ran outNoted in the build log, 23 Sep 2026.
    206Books on the shelf at the startNoted in the build log, 22 Sep 2026.
    3,392Pixels across: the screen a layout was saved onNoted in the build log, 26 Sep 2026.
    1,666Pixels across: the window it was applied toNoted in the build log, 26 Sep 2026.
    37Changes shipped on 28 SepCounted in the code history.
    20Days from the first book to a live service22 Sep to 11 Oct 2026, day 1 to day 20.
    241Commits since 25 SepCounted in the code history, 11 Oct 2026.
    50Test filesCounted, 11 Oct 2026.
    48Milliseconds a search takes (typical)Measured, 26 Sep 2026.
    64Memory of the Mac Studio (gigabytes)System report, 11 Oct 2026.