Project Summary
Company
Automotive Parts Retailer
Industry
E-commerce
Services
AI Software Development, LLM Evaluation
Location
Europe
An automotive parts retailer needed product listings for the entire Ford parts catalogue, 587,000 SKUs, each written separately for eBay, Amazon and its own webshop. We built a generation system with an automated quality gate in front of every listing and a feedback loop that improves with each human correction.
Full Catalogue Coverage: 1.76 million listings produced across three sales channels.
Automated Quality Gate: Every listing is approved, routed to a reviewer, or regenerated with specific feedback.
Continuous Improvement: Reviewer corrections are fed back into the system as examples.
The Challenge
Each marketplace has its own format, rules and expectations, and a catalogue of 587,000 parts is far beyond what a team can write by hand. Generating listings automatically solves the volume problem but creates a quality one, because a listing is only useful if it is accurate.
The client needed listings at full catalogue scale, tailored per channel, with quality controls strong enough to trust the output.
The Approach
We built a pipeline that generates listings tuned to each channel. In front of it sits an LLM evaluator that checks every listing before it goes live. A listing either passes, is routed to a human reviewer, or is retried with specific feedback on what needs to change.
Every correction a reviewer makes is captured and fed back as a few-shot example, so the system gets more accurate the more it is used. The platform is built with React, TypeScript, FastAPI and PostgreSQL on Google Cloud.
The Impact
The client now has 1.76 million listings covering the full 587,000-part catalogue across eBay, Amazon and its webshop. Human review is reserved for the listings that need it, and the system’s accuracy continues to improve with every correction.
