Umma's Kitchen: Preserving Family Recipes with AI

There's a specific kind of panic that hits when you realize you've never actually written down how your mom makes her food. She cooks by memory, by feel, by the smell of garlic hitting hot oil. You've watched her a hundred times and still couldn't recreate it if you tried. That's the problem Umma's Kitchen is built to solve.

Mom opens the app, hits record, and just talks in Korean. She rambles the way moms do: "You take some short ribs, maybe a pound, but honestly use more if you have it, and the marinade is just soy sauce, pear if you have one, garlic, you know..." The app transcribes her, translates if needed, and turns that messy beautiful monologue into a clean, organized recipe card with ingredients, steps, and her own tips preserved in both languages, with a voice memo of what the dish means to her.

How it works under the hood

I went back and forth on the architecture before landing on a simple two-model pipeline. The first step takes audio in and gives structured text out: a multilingual speech model handles transcription and translation in a single pass, so mom's voice memo in Korean comes back as both the original Korean transcript and an English translation. The second step takes text in and gives a warm recipe out: a language model structures the transcript into ingredients, steps, and tips, while preserving culturally specific terms.

The part that surprised me

The hardest part wasn't the AI. It was the tone. Default LLM output for recipes reads like a Wikipedia article — "Combine the marinade ingredients in a medium bowl." Mom doesn't talk like that. She says "mix it all together, you'll know when it looks right." Getting the model to keep that voice — to write like her and not like a content farm — took more prompt iteration than anything else.

The fix was passing both the original-language transcript and the translation to the structuring model, with explicit instructions to preserve her phrasing and her metaphors. Now when the recipe says "the broth should look like weak tea," that's her line, not the model's.

What's next

Right now, it's a working prototype. I'm trying to turn it into a real production app where families can pay for a private space to preserve recipes and voice recordings from parents and grandparents.

Next, I'm adding user accounts, family recipe pages, and support for more languages beyond Korean and English. The goal is simple: let people speak naturally in their own language and turn those memories into recipes their kids can keep forever.

Because some recipes are too important to lose.