DEV Community

Cover image for "Namak Andaaz Se": Saving Unwritten Family Recipes with Gemma, Offline
vimla
vimla

Posted on Edited on

"Namak Andaaz Se": Saving Unwritten Family Recipes with Gemma, Offline

Hacktoberfest Open-Source AI Challenge: Ghost in the Machine Submission 👻🌼

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 4: Ghost in the Machine

What I Built

Haath ka Swaad (हाथ का स्वाद, "the taste of her hands"): record someone cooking, get back a recipe card in their words, with their voice on every step.

The problem

The best food in Indian families was never written down. It lives in andaaz: "ek mutthi atta" (a fistful of flour), "namak andaaz se" (salt, by feel). Ask for the recipe and you hear "bas, dekh ke seekh lo" (just watch and learn). One day there's no one left to watch.

A 40-minute video isn't a recipe. A typed-up recipe loses her voice. Cloud apps want an account and your grandmother's voice on their servers.

The solution

Record Mummy, Nani or Dadi cooking while she talks in Hindi, Marathi, Gujarati or Hinglish. You get a card with:

  • her exact words, in her script, with an English meaning underneath
  • approximate measures beside them (~1 tsp), marked as estimates
  • a ▶ on every step that plays her voice from the moment she said it
  • "Ask next time" questions for what she skipped

All of it runs on a laptop, offline.

The ghost in the machine: long after she's gone, you press ▶ and hear "thoda sa haldi, jab tak khushboo na aaye" ("a little turmeric, until you can smell it") in the voice that taught you.

Demo

Recipe card: her words, English meaning, approximate measures

To try it without downloading any models, run haath ui --mock. It loads a sample Dal Baati recipe.

Code

GitHub logo vimla01 / recipe-ghost

Record a family elder cooking and explaining in Hindi, Marathi, Gujarati or Hinglish, and get a recipe card.

हाथ का स्वाद · Haath ka Swaad

Haath ka Swaad: every family recipe, in the cook's own words and the cook's own voice

▶ Watch the demo video

For every recipe that only exists in her hands.

Record Mummy, Nani or Dadi cooking while she explains in Hindi, Marathi, Gujarati or Hinglish ("thoda sa haldi, jab tak khushboo na aaye"), and get back a recipe card that keeps her words adds approximate measures beside them, and lets you press ▶ on any step to hear her voice explain it. Everything runs on your laptop. Nothing leaves it.


The problem

The best food in most Indian families was never written down. It lives in andaaz: "ek mutthi atta" "namak andaaz se", "jab tak tel chhod de". When you ask for the recipe you get a shrug and "bas, dekh ke seekh lo" (just watch and learn). And…

How I Built It

her voice ──► faster-whisper ──► Gemma 4 via Ollama ──► guard ──► recipe card ▶
              transcript +        her words +             drops anything
              timestamps          ~estimates (JSON)       she never said
Enter fullscreen mode Exit fullscreen mode
  1. faster-whisper (large-v3-turbo) transcribes on the laptop with timestamps and detects the language.
  2. Gemma 4 (gemma4:e4b) via Ollama turns the messy, code-switched transcript into a recipe. Ollama's format field forces the output to match my Recipe JSON schema.
  3. Guard: every quote, ingredient phrase and cue Gemma returns must appear in the transcript (Devanagari and Hinglish normalised), or it gets removed. Each step is snapped to its timestamp, so ▶ plays the right moment.
  4. Review: you tick each line together with the cook. Editing a line removes its tick.

Why Gemma

  • It runs on an ordinary laptop. e4b runs on a CPU-only i5 with 16 GB of RAM, and e2b on slower machines. Families don't have GPUs.
  • It understands Hindi, Marathi and Gujarati. Kitchen talk switches between Devanagari and Hinglish mid-sentence. Gemma keeps her phrasing instead of flattening it into English.
  • Reliable structured output. With a JSON schema through Ollama, every reply parses into a card, which the guard can then check line by line.
  • I can tune it. I turned thinking mode off for speed, chunk long recordings, and switch sizes with --model. A closed API doesn't let you do that.
  • Open weights, so it stays local. That's what makes the rest of this possible (next section).

Why Does Open Innovation Matter?

These are recordings of mothers and grandmothers in their own kitchens.

  • Private: no server ever hears her voice.
  • Offline: after a one-time download, it works with the Wi-Fi off.
  • Permanent: recipes are plain JSON and audio files, plus a book.html that plays offline. No app that can shut down or account that can be deleted.
  • Free: no subscription to set up for your parents.

The Spirit of It

If someone's food only lives in their hands, record them this week, while you still can.

Prize Categories

  • Best Use of Gemma: Gemma 4 runs locally and turns multilingual kitchen talk into a schema-checked recipe that keeps the cook's own words.

Top comments (0)