E-commerce Product Content Generator (Vision + LLM)

The Problem

A large e-commerce retailer struggled with standardizing product content. Products arrived from suppliers with inconsistent structural data and poor-quality images. Manually rewriting titles, descriptions, and extracting specifications for tens of thousands of SKUs was a massive bottleneck.

The Implementation

I built a dual-pipeline system combining Vision AI and Language Models. The pipeline takes raw product images and basic supplier data as input.

First, the Vision model analyzes the images to extract visual attributes (color, pattern, material, style). Then, an LLM combines this structured visual data with the supplier text to generate SEO-optimized titles, compelling descriptions, and standardized feature lists.

• Vision model fine-tuned for e-commerce specific feature extraction.
• LLM pipeline with strict output formatting to match database schemas.
• Asynchronous processing queue to handle large batch uploads.
• Automated quality checks and fallback mechanisms for low-confidence generations.

Outcome

The system successfully generated content for over 50,000 products in production. It reduced the time-to-market for new catalog items by a factor of 10 and established a consistent brand voice across all product listings.