Building an ETL Pipeline for 15,000+ Luxury Items

The Problem

When you're processing thousands of luxury items for eBay, manual categorization becomes the bottleneck. Each item needs:

Correct category classification (suits, sport coats, outerwear, etc.)

Measurement extraction from unstructured text

Brand profiling for marketing copy

Templated HTML generation

The Solution

I built an ETL pipeline with three core components:

1. Category Classification

Using sentence-transformer embeddings, items are classified into 60+ categories. When embeddings don't match with high confidence, RapidFuzz provides fuzzy matching fallback.

2. Measurement Parsing

20+ regex parsers extract measurements from free-text descriptions, normalizing them into structured JSON.

3. GPT Integration

For brand profiling, I integrated OpenAI's API with JSON caching to minimize costs while generating dynamic marketing copy.

Results

15,000+ items processed

5× throughput improvement

$250K+ annual revenue enabled

The key insight: AI doesn't replace human judgment—it handles the 95% of routine cases so humans can focus on edge cases.