Beyond Simple OCR: Building an Autonomous VLM Auditor for E-Commerce Scale

In the world of global e-commerce, “dirty data” is a multi-billion dollar problem. Product dimensions (Length, Width, Height) are often inconsistent across databases, leading to shipping errors, warehouse mismatches, and customer returns. Traditional OCR struggles with complex specification badges, and manual auditing is impossible at the scale of millions of ASINs. Enter the Autonomous VLM Auditor — a high-efficiency pipeline utilizing the newly released Qwen2.5-VL to extract, verify, and self-correct product metadata. The Novelty: What Makes This Different? Most Vision-Language Model (VLM) implementations focus on captioning or chat. This project introduces three specific technical novelties: 1. The “Big Brain, Small Footprint” Strategy To process over 6,000 images at scale, we utilized 4-Bit Quantization (NF4) via BitsAndBytes. In the world of VLMs, memory is the primary bottleneck. By compressing the model's weights from 16-bit to 4-bit, we reduced the VRAM footprint by nearly 70%.