What I Built

I built Dadi's Rasoi (Grandma's Kitchen), a warm, local-first web application that turns scattered voice notes into a structured, searchable family recipe book.

I built this for my grandmother (Dadi) and my broader family. Like many families, our traditional recipes exist entirely in someone's memory or in rambling Hindi/Hinglish voice notes sent over WhatsApp. Dadi's Rasoi solves this by taking her raw voice recordings, automatically transcribing them, organizing the ingredients and instructions, and saving them into a beautiful digital cookbook. Most importantly, it is specifically prompted to preserve her authentic phrasing (like "namak apne hisaab se" / salt to taste) and her personal cooking tips, rather than hallucinating generic measurements.

Code

shivamgravity / hf26-recipe-archived

Dadi's Rasoi ❤️

Preserve the recipes that live in someone's memory.

Built for Hacktoberfest 2026 — Weekend Challenge: Build for a Friend

💡 The Problem

Family recipes frequently exist only in someone's memory or scattered voice messages.

🥘 The Solution

Dadi's Rasoi is a small, warm personal web application that converts those voice explanations into structured, searchable family recipes.

🛠️ Local AI Architecture

Once the models are installed, the core AI pipeline can run without sending family voice recordings to a remote AI provider.

Voice
  ↓
faster-whisper (Local Speech Recognition)
  ↓
Transcript
  ↓
Gemma via Ollama (Local LLM)
  ↓
Structured Recipe
  ↓
Pydantic validation
  ↓
SQLite (Local Cookbook)

🤔 Why Open AI?

Privacy: Keep family data under your control. The local models ensure nothing leaves your laptop.
Model Flexibility: Swap out the core extraction model as open-weight models evolve.
No Proprietary Lock-in: Ensure the family cookbook isn't dependent on expensive…

View on GitHub

How I Built It

The application is written in Python using Streamlit to create a warm, minimal, and non-intimidating UI. The AI pipeline is built entirely around open-source and open-weight AI running 100% locally on consumer hardware:

Local Speech Recognition (faster-whisper): The app takes the audio file and processes it directly on the CPU using faster-whisper (the small model with int8 quantization). It naturally detects and transcribes Hindi and Hinglish without aggressively forcing translations, preserving her exact words.

Local Structured Extraction (Gemma 3): The transcript is sent to Google's open-weight Gemma 3 (4B) model, served locally via Ollama. Gemma is prompted to act as a strict data extractor, parsing the raw text into a structured JSON schema (validated by Pydantic) containing titles, ingredients, steps, and a special section for "Family Tips".

Local Storage (SQLite): The user reviews the AI's extraction, makes any necessary manual edits, and saves the final recipe card to a local SQLite database for searching and browsing.

Why Does Open Innovation Matter?

Open innovation is the entire reason this project works for this specific use case. Voice recordings of your family members are deeply personal. Sending a grandmother's voice memos to a closed, third-party API server to be processed (and potentially trained on) felt completely wrong.

By using open-weight models like Gemma 3 and open speech frameworks like Whisper running natively on my own device, Dadi's Rasoi is 100% private. The audio and the recipes never leave the laptop. Furthermore, because these models are open, I was able to utilize lightweight, quantized versions that execute comfortably on a consumer laptop without racking up cloud API bills. Open AI gave me the freedom to keep family memories entirely within the family.

DEV Community

My Dadi's Recipes Were Trapped in Voice Notes, So I Built Her an AI Cookbook