How the AI works
Careful reading, done by machines, checked like a proofreader would.
Kitchenlore uses modern vision, speech and language models to do what a patient grandchild does at the kitchen table. Here is exactly how, in plain terms, and the things we will not let the models do.
i. The pipeline
Five steps between the card and the page.
- 01
Read the card
Vision + languageA photo of the card goes to a vision-capable language model with our house rules. It transcribes everything legible line by line, margins and crossings-out included, and proposes a structured recipe in the same pass.
- 02
Hear the memo
Speech to textA voice memo goes to a speech-to-text model that returns a verbatim transcript in the language she spoke: the pauses, the asides, the “no, wait, first the mustard seeds.” The transcript is kept whole and attached to the recipe. Only then is it structured.
- 03
Structure the recipe
Language modelText from a card, a transcript or a pasted chat is turned into a recipe under strict house rules: never invent, keep amounts exactly as written, keep the writer's wording in the steps, keep the original language. The output must match a fixed recipe schema, validated with zod. Anything malformed is rejected and the request tried again, never patched by hand.
- 04
Keep their words
Deterministic checkEvery “In her words” quote the model proposes is compared against the original source, word for word, ignoring only case, curly quotes and spacing. A quote that is not in the source is dropped. Not reworded, not softened: dropped. This check is ordinary code, not a model, so it cannot be talked into anything.
- 05
Mark, never guess
House rule + checkIf a word cannot be read or heard with confidence, the model is told to write [hard to read] in place. We also normalize every other way a model might say it (“illegible”, “???”) into that one honest mark, and list each spot under the recipe for the family to fill.
ii. Routing
Model-agnostic, on purpose.
We do not bet the family recipes on one company. Reading, hearing and structuring are routed independently across leading model providers, today OpenAI and Anthropic. Each job has a primary route and a fallback. If a provider is slow, down, or returns something that fails our checks, the same request goes to the next route automatically.
Because every route runs the same house rules and every answer passes the same schema and quote checks, the result you get does not depend on which model happened to answer. When a better model for reading cursive appears, we can switch to it without changing a single rule.
And when no model is available at all, pasted text still works. A small structurer that runs on our own servers, with no model behind it, splits ingredient lines from steps using plain rules. It is less clever, and it says so, but it never invents anything either.
iii. The rules
What we never let the models do.
- Invent. No new ingredients, amounts, steps, times, temperatures, names, dates, stories or feelings. If the source does not say it, the field stays empty.
- Guess at a smudge. Even when a guess looks obvious. The family fills gaps, signed with their names.
- Polish a quote. Her words are copied, never paraphrased, and checked against the source.
- Translate unasked. A card in Norwegian, or in Hindi and English, stays as written. A short note says which languages it holds.
- Train on your recipes. Family recipes are never used to train our models or anyone else’s. We use providers through their APIs, under terms that do not use API data for training.
- Copy a voice. We transcribe memos. We will never clone or synthesize a voice.
Cook’s note
A note on the method
Built AI-native.
For fifty years the recipe box had no good answer. Scanning kept the picture but lost the recipe. Typing it out kept the recipe but lost the voice. Neither could read Mormor’s cursive or sit through forty minutes of Dadi explaining dal.
Kitchenlore exists because models can now read and listen well enough to do this work. We built the product around that from the first day: the models do the reading, and plain, testable code does the checking. The models are good at reading. The checks make sure they stay honest.
Seven people, building AI full time, in Minneapolis
iv. Limits
Where it falls short.
Models still miss things. A very faded pencil card may come back with more [hard to read] marks than you would like, and a memo recorded next to a running fan will lose words. We would rather show you a gap than hand you a confident mistake, so expect gaps on hard sources. Better photos help more than anything; our help centre has a short guide.