What I Built

I built Dadi's Probashi Kitchen, a local AI companion designed for my friend Nafis who recently moved from Bangladesh to Melbourne, Australia for his studies.

Living abroad for the first time, Nafis faced severe homesickness and missed traditional homemade comfort food like Kacchi Biryani and Shorshe Ilish. He struggled with two major issues:

Standard Aussie supermarkets don't carry specific local Bangladeshi ingredients (like Ilish fish or local spices).

Family recipes passed down from grandmothers use vague measurements like "ek mutho" (a handful) or " আন্দাজমতো " (to taste), which are impossible for a beginner cook in a new country to decode.

Dadi's Probashi Kitchen acts as a warm, loving Bangladeshi grandmother ("Dadi"). Nafis can type any traditional dish he craves, and Dadi comforts him in "Banglish" (Bengali-English), suggests standard supermarket alternatives available in Western stores (e.g., swapping Ilish for salmon or sea bass), and converts vague traditional measurements into precise cups and grams.

Demo

Live App (Render): https://probashi-recipe.onrender.com/ (Note: Hosted on a free-tier server which occasionally runs low on RAM for local LLM extraction. Please see the video demo for local CPU execution!)

When Nafis tested it, he said: "It actually sounds like my Dadi! Replacing the hard-to-find mustard fish with sea bass made making dinner after lectures so much less intimidating."

Code

ZuhairHossain / probashi-recipe

👵🏽 Dadi's Probashi Kitchen

A locally-run AI assistant built for the Hacktoberfest 2026 DEV Weekend Challenge: Build for a Friend.

The Story

My friend Nafis recently moved from Bangladesh to Melbourne, Australia. He has been incredibly homesick for authentic Bangladeshi food like Kacchi Biryani and Shorshe Ilish, but he struggles to find local ingredients in standard Aussie supermarkets and doesn't understand traditional vague measurements (like "ek mutho" / a handful).

This project solves that problem. "Dadi's Probashi Kitchen" acts as a warm, loving Bangladeshi grandmother. It takes traditional recipe requests, comforts the user in "Banglish," swaps hard-to-find ingredients for Western supermarket alternatives, and converts measurements into exact grams and cups.

Why Open Source?

Cultural Customization: Using an open-weight model allowed me to heavily customize the system prompt to capture the exact warmth, vocabulary, and specific tone of a real South Asian grandmother, avoiding the robotic feel of closed…

View on GitHub

How I Built It

The application is built with Python and Streamlit for the user interface, powered entirely by open-source AI:

Model: Google's Gemma 2B (gemma2:2b), an open-weight model selected for its lightweight efficiency and strong instruction-following capabilities.

Inference: Ollama runs the model locally via CPU inference without requiring paid API tokens or a dedicated GPU.

Containerization: The app is fully Dockerized (Dockerfile + start.sh) for deployment options on cloud platforms like Render.

# System prompt snippet defining Dadi's persona in Streamlit
system_prompt = """You are a warm, loving Bangladeshi Dadi (Grandma). Your grandchild just moved abroad and misses your cooking.
1. Comfort them affectionately in a mix of English and Banglish (use sweet words like 'shona', 'bhaiya', 'dadibhai').
2. Adapt recipes using standard Western supermarket ingredients.
3. Convert vague measurements like 'ek mutho' (a handful) into standard cups, grams, or tablespoons."""

Why Does Open Innovation Matter?

Open innovation made two crucial aspects of this project possible that proprietary closed APIs could not deliver:

Cultural Persona Customization without Corporate Censorship or Robotic Tone: Closed commercial APIs often enforce rigid, corporate conversational styles or fail to adopt culturally nuanced dialects like "Banglish". Open-weight models like Gemma 2B allow complete control over system prompts and temperature settings, producing an authentic, comforting persona.

Offline Execution & Complete Recipe Privacy: Family recipes are personal heritage. Running Gemma 2B locally via Ollama means recipes and personal inputs never get sent to cloud logging servers or mined for advertising. Furthermore, because Gemma 2B runs locally on standard laptop CPUs, my friend can run the app offline in his apartment even when his internet goes down—costing $0 to host or maintain indefinitely.

DEV Community

Dadi's Probashi Kitchen: Bringing Home to Expat Friends with Open-Source AI