---
title: Dukaan Saathi – AI WhatsApp Reply Assistant
emoji: 🚀
colorFrom: yellow
colorTo: yellow
sdk: gradio
sdk_version: 6.18.0
python_version: '3.13'
app_file: app.py
pinned: false
---

# 🏪 Dukaan Saathi — AI WhatsApp Reply Assistant for Indian Shop Owners

> *Built for the [Gradio × Hugging Face Build-Small Hackathon 2026](https://huggingface.co/build-small-hackathon) · Track: Backyard AI*

## The Problem

My neighbor runs a medical store in Ahmedabad. Every day, 40–60 customers message him on WhatsApp asking about medicine availability, prices, and home delivery. He's one person behind a counter — he can't reply to everyone fast enough, and he loses orders because of it.

Big AI tools like ChatGPT aren't an option for him: he doesn't want customer prescription details going to some American server. And he doesn't have time to learn a new app.

**Dukaan Saathi** solves this with a small, privacy-respecting model that runs entirely on Hugging Face infrastructure — no data sent to big cloud APIs, no subscriptions, no complexity.

## What It Does

1. Shop owner sets up their profile once (shop name, type, items they sell)
2. Paste any customer WhatsApp message
3. Choose language: Hindi, Gujarati, English, or Hinglish
4. Get a perfectly worded, human-sounding reply in 3 seconds
5. Copy → paste → send

## Why Small Models Are the *Right* Choice Here

This isn't a compromise — small models are actually **better** for this use case:

- **Privacy**: Customer messages (some containing prescription details or personal info) never leave a controlled environment
- **Speed**: 7B models respond in seconds, not minutes
- **Cost**: Shop owners can't pay per-API-call; small models make this sustainable
- **Offline potential**: The same model can run locally on a mid-range laptop

## Model

**Qwen/Qwen2.5-7B-Instruct** — 7 billion parameters (well within the 32B cap)

Chosen because:
- Best-in-class multilingual performance for Hindi, Gujarati, and Hinglish
- Instruction-following quality that produces natural, conversational replies
- Lightweight enough to be genuinely "small model" in spirit

## How to Run Locally

```bash
git clone https://huggingface.co/spaces/YOUR_USERNAME/dukaan-saathi
cd dukaan-saathi
pip install -r requirements.txt
HF_TOKEN=your_token python app.py
```

## Hackathon Info

- **Event**: [Gradio × HF Build-Small Hackathon 2026](https://huggingface.co/build-small-hackathon)
- **Track**: Chapter One — Backyard AI
- **Constraint**: ≤32B parameters (using 7B)
- **Builder**: Parth Bhuptani ([@ParthBhuptani](https://github.com/ParthBhuptani))