ListSmart: Turn Product Photos into E-commerce Listings with AI

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

What I Built

ListSmart converts product photos into ready-to-publish Meesho and Etsy listings, including titles, descriptions, five bullet points, search tags, and suggested prices.

It helps small sellers save time, write better listings, and price products more confidently.

Key features:

📸 Product detail extraction using Gemma.
✍️ Listing generation using fine-tuned Qwen3-8B.
💰 Competitor-based pricing using SerpApi.
🔒 Local inference for privacy and zero per-listing API costs.
✏️ Editable listings with character counters and one-click copy.

The proposed business model is 20 free listings, followed by approximately ₹199/month. Pricing is still being validated.

Demo and Code

GitHub: https://github.com/Amitrawal1/ListSmart

Run locally:

git clone https://github.com/Amitrawal1/ListSmart.git cd ListSmart/backend python3 -m venv .venv .venv/bin/pip install -r requirements.txt .venv/bin/uvicorn app.main:app --port 8001

Open http://localhost:8001. Add ?demo=1 to try sample data.

How I Built It

The project uses Gemma for vision, Qwen3-8B with LoRA fine-tuning for listing generation, Tinker for training, and SerpApi for competitor pricing.

The stack includes FastAPI, Ollama, and a lightweight HTML/CSS/JavaScript frontend. The repository also includes training, evaluation, and setup scripts.

What I Learned from Version 1

My first fine-tune reduced validation loss from 2.61 to 1.48, but performed worse than the base model on important quality metrics.

The blind evaluation preferred the base model 77.5% of the time, compared with just 0.8% for v1.

The main problems were repetitive keyword stuffing and invented product details.

For v2, I'm cleaning the training data, removing unsupported claims, and enforcing platform rules. The final evaluation is still pending, and I'll share the results once measured.

Why Open Source Matters

Open models let me inspect training data, fix model behaviour, and run inference locally. Sellers can keep product photos on their own devices, while local inference avoids per-listing API charges.

The project also reduces dependence on a single model provider and makes the system easier to modify.