Volume versus interface latency
Vast amounts of product data had to be processed and the result shown in the UI within a time users would accept — the constraint that ruled out several architectures.
Mapping categories, features and feature values between taxonomies is where product content onboarding slows to a crawl. Bintime integrated an embedded OpenAI model into Gepard’s PIM & Syndication platform so the system proposes the mapping and the user confirms it, with 75% auto-mapping accuracy.
Gepard PIM
Visit the client’s pageGepard is a PIM platform that brings automation into product information management for e-commerce businesses. Its PIM & Syndication platform collects, manages, enriches and distributes product data in the format each sales channel requires. Brands exchange product marketing content freely; retailers onboard and adapt it in an automated way.
The solution delivers more than 120 million product descriptions per month across multiple retail platforms. At that volume, any step that still needs a human decision per attribute becomes the constraint on the entire pipeline.
Three constraints shaped every decision: the volume of product data, the response time users would tolerate in the interface, and the need to keep simple mappings simple while leaving room for AI-powered suggestions on the complex ones.
Vast amounts of product data had to be processed and the result shown in the UI within a time users would accept — the constraint that ruled out several architectures.
Simple mappings had to stay fast and deterministic, while the same system had to accommodate AI-powered suggestions for the complex cases.
Reliability and scalability had to hold as new taxonomies and products arrived, with no additional manual work per source.
The project began with a discovery phase to understand the specific requirements and hurdles of e-commerce operations. Only then did the team engineer an architecture designed to fit the existing platform infrastructure, integrate the embedded OpenAI model, and build the intelligent mapping system with a responsive interface. QA testing verified reliability and mapping precision, and continuous updates carried market feedback back into the feature.
For adaptive planning and continuous improvement of the MVP features.
To ensure modularity and ease of maintenance.
For streamlined and automated code deployment.
To make the platform intuitive and easy to use.
To protect data integrity and privacy.
Cover data mapping, the most complex process in product content syndication and enrichment, instead of automating the parts that were already cheap.
Category mapping, feature mapping and feature-value mapping were automated alongside a table-based import system for e-commerce businesses.
Fields arrive already recognised, so taxonomy mapping starts from a proposal rather than a blank table.
Values are matched against the content taxonomy without a person searching for each one.
Where a match is uncertain, the system proposes candidates instead of failing silently.
Suggestions cover the structural level, not just individual values.
Confirmed matches are applied in a single action, which is where most of the saved time comes from.
Architecture that fits the existing platform infrastructure.
Code development and UI integration of the mapping system.
Focused on the embedded OpenAI model integration.
Oversees the project timeline and deliverables.
Tests the accuracy and reliability of AI-powered mapping.
Suggestion accuracy hit 75%, letting brands and retailers manage product information more effectively.
Accurate and automated across all three levels, reducing manual effort and errors.
Mapping extends across various retail platforms without proportional manual work.
Easy navigation and operation of the mapping functions, at the volume the platform handles.
The embedded OpenAI model now covers data mapping inside Gepard’s PIM & Syndication platform — the most complex process in product content syndication and enrichment. Automating the step everyone else leaves manual is what changes the economics of onboarding a new catalogue.
Recognition of import file fields, mapping of feature values, value suggestions, AI-generated category and feature mapping, and one-click mapping shortcuts.
Auto-mapping suggestion accuracy reached 75%. The rest stays a human decision, with candidates proposed rather than guessed.
An embedded OpenAI model recognises categories, features and feature values, supports new taxonomies and improves accuracy over time.
Seven people: a solutions architect, two backend and one frontend developer, an AI specialist, a project manager and a QA engineer.
Tell us which process eats the most human time. We will come back with what an AI-assisted version would take: discovery scope, architecture questions to answer first, and where accuracy would realistically land.
We reply within one business day.
We will get back to you within one business day to discuss the details.