Behind the scenes

How we taught AI to actually read a recipe

Parsing a recipe sounds simple. Doing it reliably across thousands of messy formats is the hard, interesting part.

March 30, 2026 · 6 min read
How we taught AI to actually read a recipe

A recipe is messier than it looks

“2 Tbsp butter” is easy. But recipes hide quantities in prose, split steps across paragraphs, bury the yield in a headnote, and assume context a human just knows. Teaching software to understand all of that took real work.

Quantities, substitutions and scaling all have to survive the trip from messy text to clean card.
Quantities, substitutions and scaling all have to survive the trip from messy text to clean card.

Structure first, then judgement

We combine layout understanding with language models so Kitchin knows the difference between an ingredient, an instruction, and a story about someone's trip to Italy. Then it normalises units, infers servings, and estimates nutrition.

The goal: a card that's faithful to the original, but easier to cook from.
The goal: a card that's faithful to the original, but easier to cook from.

It's not perfect — and that's why every card stays fully editable. But it's right often enough to feel like a little kitchen magic.


Back to all posts