I Built a Recipe Book for My Dadi, Using AI That Never Leaves My Laptop

My dadi's best recipes have never been written down.

Ask her how much salt goes in, and she says "ek mutthi." Ask how much ghee, and it's "thoda sa." Ask how long to cook it, and the answer is "jab tak khushboo aaye." Her recipes live in her hands and her memory, and one day they will be gone with her.

I wanted to save them. But recording her and uploading the audio to a cloud AI felt wrong. These are my family's private voices, on a server I don't control. And most cloud models flatten her Hinglish into stiff, robotic steps, losing the little sayings and stories that make the food hers.

So for the "Build for a Friend" challenge, I built something for her.

What I Built

Dadi ki Rasoi turns her voice notes into a printable family recipe book. Everything runs on my own laptop.

1. She talks while cooking, and I record it.
2. faster-whisper transcribes the Hindi/Hinglish audio locally.
3. A local LLM running on Ollama (Qwen and Llama) turns the transcript into a structured recipe.
4. Desi measurements become real quantities. "Ek mutthi namak" becomes an approximate weight in grams, but her original phrase always stays right beside it.
5. Her sayings and stories are kept word for word in a "Dadi ki baat" box on each recipe.
6. Everything exports as a printable PDF book with a cover, an index, and one recipe per page, with Devanagari font support.

Code: https://github.com/vidishagupta/Dadi-ki-Rasoi
Live demo: https://web-omega-lac-94.vercel.app/
*Video link: * https://youtu.be/ykddENDuHm4?si=EDGMLkfrp-iUk2Rd

A note on the live demo: Whisper and Ollama are heavy models that can't run on serverless hosting. So the hosted version runs in Demo Mode with saved results from the real pipeline. The full offline AI pipeline runs locally, and the Decision.md has the setup steps.

Why Open Source Mattered

For this project, open source is not a bonus. It is the whole point.

Three things made the difference:

• Privacy: A family member's voice is personal. With local Whisper and Ollama, there is no upload step at all.
• Offline: The entire pipeline works with Wi-Fi turned off, so anyone can verify it in airplane mode.
• Model swap: The app has a compare view that runs the same transcript through Qwen and Llama side by side. Different models understand "ek mutthi" and "thoda sa" differently, and with open models I can see that and choose, instead of trusting a black box.

I also wrote my own desi measurement lookup table (mutthi, chutki, katori and more) and fed it to the model, so conversions stay consistent. That kind of control is hard to get from a closed API.

What I Learned

Honesty beats confidence. Dadi's measurements are estimates by nature. So the app marks low-confidence ingredients, and the recipe editor shows the original transcript next to the structured recipe, so a human can fix anything before printing. The AI is never allowed to invent an ingredient or a step she did not say.

Keep her voice, not just her steps. A recipe is not only quantities. Keeping her own words and small stories in the book is what makes it hers and not a generic recipe card.

Know what the hosting can do. Heavy local models don't fit on serverless platforms, so I built a Demo Mode for the live link and kept the real pipeline local-first, which fits the whole idea anyway.

Limitations: Whisper can mishear regional words or noisy kitchens, smaller local models can miss a step, and gram conversions are approximate. That is why everything stays editable before it is printed.

The Real Test

This project was built for one real person. The goal was never a perfect demo. It is a printed book in my dadi's hands, with her recipes in her own words, that her grandchildren can cook from long after today.

What's Next

More regional languages, better handling of noisy kitchens, and a photo of each dish on its recipe page.

The project is open source and Hacktoberfest-friendly. If your family has recipes that only live in someone's head, fork it and capture them before they are lost.

GitHub: https://github.com/vidishagupta/Dadi-ki-Rasoi
This is a submission for the Hacktoberfest Weekend Challenge: Build for a Friend

What I Built

Demo

Code

How I Built It

Why Does Open Innovation Matter?

My Agent Session

Prize Categories