DukaanWriter: WhatsApp-Ready Product Listings in English and Hindi

Pranjal Aman — Tue, 13 Oct 2026 08:38:34 +0000

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle

What I Built

DukaanWriter turns a product photo, a name and a price into a catalog listing in English and Hindi, formatted so it can be pasted straight into WhatsApp or Instagram.

The person who would pay for it is a small shopkeeper or home seller (clothes, jewellery, sweets, gifts) who sells through WhatsApp Business and Instagram. They post new products constantly, and every post needs a clear title, a short description and hashtags. Writing that in two languages, for every product, is slow, and most AI writing tools charge per use or by subscription. That is a hard sell for someone whose whole stock is a few hundred items.

You add one photo, the product name, the price in rupees, and an optional note (size, material, colour). The app returns two cards, English and Hindi, each with:

a title in WhatsApp bold,
a short description,
the price line,
hashtags.

Each card shows a WhatsApp-style preview of exactly what will be sent. You can edit any field before you send it. Copy puts the text on the clipboard, and Share on WhatsApp opens a wa.me link with the text filled in, so there is no WhatsApp API key and no cost.

Demo

Code

pranjalaman / DukaanWriter

DukaanWriter

A local tool that turns a product photo, name, and price into a catalog listing in English and Hindi, ready to paste into WhatsApp or Instagram.

Product photos and prices are sent only to Ollama on your own computer. After the model is downloaded, a listing does not need the internet.

What you need

Windows, macOS, or Linux
Python 3.11 or newer
Ollama
About 16 GB of RAM is comfortable. A 4B vision model can run on CPU with 8 GB, slowly.

Setup

Install Ollama and start it.
Pull a vision model. Check the exact tag with ollama list or the Ollama library if this one has moved:

ollama pull gemma3:4b

If Hindi comes out stiff, pull the fallback and type its name under Model at the top right:

ollama pull qwen2.5vl:3b

Install the app and open it:

py -3 -m venv .venv
.venv\Scripts\activate
pip

…

View on GitHub

How I Built It

Inference: Ollama running a small open-weight vision model on the user's own machine. gemma3:4b is the default, and qwen2.5vl:3b is the fallback if the Hindi comes out stiff.
App: Streamlit for the interface, Pillow to shrink the photo to 512 px on the long side before it goes to the model. The code is two files, app.py for the UI and formatting and generate.py for the prompt and model call. generate.py can also be run straight from a terminal.
Structured output: the model is asked for strict JSON with an English and a Hindi block. If the reply is not valid JSON, the app retries once, and if that fails it shows the raw text in an edit box instead of crashing.
The model never writes the price. It is told not to put a price in the description, and the app adds the price line itself. A small model should not be trusted with the one number a customer will actually act on.
Hashtags are English on both cards, so Instagram and WhatsApp search still match them.
Anti-hallucination rule: the prompt tells the model to describe only what is visible in the photo or stated in the note, so it does not invent a fabric, size or brand.

Run DukaanWriter on your own computer

DukaanWriter runs fully on your machine. The vision model runs locally through Ollama, so product photos and prices never leave your laptop, and there is no per-listing cost.

What you need

Windows, macOS, or Linux
Python 3.11 or newer
About 4 GB of free disk space for the model
8 GB RAM works (slowly, on CPU). 16 GB is comfortable.

1. Install Ollama and download the model

Install Ollama from ollama.com/download. On Windows and macOS it starts in the background after install. On Linux:

curl -fsSL https://ollama.com/install.sh | sh

Then download the vision model (about 3.3 GB, one time):

ollama pull gemma3:4b

Optional smaller fallback, if Hindi feels stiff or your machine is slow:

ollama pull qwen2.5vl:3b

Model tags change sometimes. Check with ollama list.

2. Get the code

git clone https://github.com/pranjalaman/dukaanwriter.git
cd dukaanwriter

3. Install the Python packages

Windows (PowerShell):

py -3 -m venv .venv
.venv\Scripts\activate
pip install -r requirements.txt

macOS / Linux:

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

If PowerShell blocks activate, skip it and use .venv\Scripts\python -m pip install -r requirements.txt, then .venv\Scripts\python -m streamlit run app.py in step 4.

4. Start the app

streamlit run app.py

Your browser opens at http://localhost:8501. The chip at the top should say Model ready · gemma3:4b. If it says Ollama not running, open the Ollama app and refresh the page.

5. Make a listing

Upload one product photo.
Type the product name and the price in rupees.
Optional: add a note with the size, material, or colour. The model only mentions these if you write them here or they are visible in the photo.
Click Make listing / लिस्टिंग बनाएँ.

You get two cards, English and Hindi, each with a WhatsApp-style preview. Edit any line, then use Copy or WhatsApp पर भेजें. The price line is added by the app, never by the model.

On my laptop a listing took about 5 seconds.

Use it from your phone

Keep the app running on your laptop, and connect your phone to the same Wi-Fi. Open the Network URL that Streamlit prints in the terminal (something like http://192.168.x.x:8501). The model still runs on the laptop.

If Windows asks about the firewall, allow access on private networks. Some phone browsers block Copy on a plain http:// address. If that happens, use the WhatsApp button, or long-press the text to copy it.

Check that it works offline

After the model is downloaded, turn Wi-Fi off and make a listing. It should still work.

Optional: command line, no UI

python generate.py samples/kurta.png "Red cotton kurta" --note "free size"

This prints the English and Hindi listing as JSON, plus how many seconds it took.

Settings

Change the model: click Model at the top right and type another name, or set DUKAAN_MODEL=qwen2.5vl:3b before starting.
Ollama on another machine: set OLLAMA_HOST=http://<that-machine-ip>:11434. Photos are then sent to that machine.

If something goes wrong

What you see	What to do
"Ollama not running"	Open the Ollama app (or run `ollama serve`), then refresh.
"Run: ollama pull gemma3:4b"	The model isn't downloaded yet. Run that command.
Very slow	It's running on CPU. Try `qwen2.5vl:3b`, or use a smaller photo.
Raw text instead of two cards	The model's reply wasn't valid JSON, even after a retry. Click Make listing again.
Hindi sounds translated word for word	Switch the model, and edit the Hindi card before you copy.

Why Does Open Innovation Matter?

Cost. After the one-time model download, a listing costs nothing but electricity. A seller with 200 products pays the same as one with 20, which is the only pricing that works for this customer.

Privacy. Product photos and prices, including stock that has not been announced yet, are sent only to Ollama on the same computer. After the model is downloaded, generating a listing does not need the internet at all.

Swappable. The model is one setting, either a field in the app or the DUKAAN_MODEL environment variable. If a better small model with better Hindi comes out next month, nothing else in the project changes. With a closed API I would be tied to whatever that vendor decides to offer and charge.

An honest trade-off. Small local models are weaker at natural Hindi than the biggest closed models, and that is the main risk of this project. That is why the cards are editable before you send them, why there is a fallback model, and why I treat Hindi quality as something to check on real products rather than assume. For this customer, a good-enough listing that is free, private and works offline beats a perfect one that costs per use.

What's Next

This is v1. It does not yet do bulk upload from a folder and a CSV, example listings per shop type to improve tone, or other languages like Bengali, Marathi and Tamil. Fine-tuning on a seller's own past posts so the listings sound like them is the idea I'm most interested in.

DEV Community: Pranjal Aman

DukaanWriter: WhatsApp-Ready Product Listings in English and Hindi

What I Built

Demo

Code

pranjalaman / DukaanWriter

DukaanWriter

What you need

Setup

How I Built It

Run DukaanWriter on your own computer

What you need

1. Install Ollama and download the model

2. Get the code

3. Install the Python packages

4. Start the app

5. Make a listing

Use it from your phone

Check that it works offline

Optional: command line, no UI

Settings

If something goes wrong

Why Does Open Innovation Matter?

What's Next