Snaplist: A Local AI Listing Writer for Etsy & eBay Sellers (Gemma 3 + Ollama)

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

What I Built

Snaplist is a listing writer for solo sellers on Etsy and eBay. You drop in a few photos of an item, pick a marketplace and a tone, add any facts the photos can't show (condition, size, backstory), and it drafts a title, a description, and up to 13 search tags. Every word is editable before you copy it out.

Who would pay for it? Anyone who lists secondhand or small-batch items and keeps rewriting the same kind of description. The idea is a one-time desktop package rather than a monthly subscription. My working guess is around ₹999, but that is a pricing hypothesis I haven't tested, not revenue.

Demo

Video walkthrough, generating a listing from photos of a kitchen timer: https://youtu.be/j0H6ZO6grRg

Code

Public repo, MIT licensed: https://github.com/anshraj3060/snaplist

How I Built It

The core is Gemma 3 (4B), an open-weight vision model, served by Ollama on my own MacBook Air. The app is a small Node.js server plus a browser UI. When you press Generate, the server sends your photos (base64) to Ollama's /api/chat endpoint and asks for structured JSON back, which the UI turns into an editable title, description, and tags.

I built the app with an AI coding agent (Arena) from a prompt, then ran it myself: installed Ollama, pulled gemma3:4b, switched Snaplist from its labeled demo mode to "Use my local model", and tested it with my own photos.

A few details I like:

No silent fallback. If the model fails, you see an error, never fake sample text passed off as AI output.
Demo mode is labeled "not AI generated" so nobody mistakes it for real analysis.
The model is swappable. OLLAMA_MODEL=gemma3:12b npm start is all it takes to try a bigger one.
Tags are capped at 13 and titles are limited per marketplace in code, not left to the model.

Why Does Open Innovation Matter?

Cost structure. With open weights running locally, there is no per-listing API bill. That is what makes a one-time license even thinkable. A closed API would push me toward a subscription just to cover inference.
Privacy. In local mode, photos go from the browser to the Snaplist server to Ollama, all on the seller's own computer. Unlisted products and supplier photos never reach a third party.
Control. I can inspect the prompts, change the model with one environment variable, and later fine-tune for a niche. With a closed API I get whatever the vendor ships this month.
Works offline. After the one-time model download, generation doesn't need the internet.

Where it did not beat a closed model: a 4B model is small. In my first test I uploaded photos of four unrelated items, and it merged them into one confused listing. It did read a brand name off a bottle, which was a nice surprise. Descriptions come out short and need a human edit. CPU inference can be slow, and there is a 180-second timeout. Hardware and electricity still cost something, and Gemma is open-weight under its own terms, not OSI-approved open source.

I also deliberately did not build price suggestions. Snaplist has no comparable-sales data, and I'd rather ship nothing than invent numbers.

My First User

I haven't had anyone else try it yet, so there is no first-user story to tell. Next step: put it in front of a real seller and ask what they'd pay, and what they'd want it to do next.