Back to all articles

AI-Powered Catalog Enrichment Tool: How We Made It

Info Setronica September 16th, 2025

Many businesses struggle with “empty” catalogs when trying to scale. These bare-bones product listings lack the rich details customers need to make informed purchasing decisions. 

That’s why we created SmartCatalog – an AI tool that transforms basic product information into comprehensive, enriched catalogs.

With a team of just two people and a three-week timeline, we built an AI-powered tool that goes beyond what generic AI solutions can offer. In this article, we’ll share how we built SmartCatalog, the challenges we faced, and where we plan to take it next.

Do they really need an AI-based catalog tool?

This wasn’t a hidden problem requiring extensive market research. It was right there in plain sight. Many individuals and entire companies, including our clients, struggle with catalogs that weren’t built for scaling.

It’s quite normal not to think about it at the very first step. It’s absolutely common, for example, if you start with a limited number of products in one category. But the list grows, as well as the requirements for your catalog.

Here’s what typically happens: businesses store their product data in Excel files, which become problematic when they try to upload them to fulfillment services or marketplaces. Compatibility issues arise everywhere. Some entries lack crucial attributes, others don’t have standardized SKUs, and the list of potential problems goes on and on.

People have to deal with manual fixes and workarounds, and the suffering begins. That’s exactly why creating a service that could take these original catalogs and, using AI, transform them into more versatile product management tools became a priority.

The question wasn’t if people needed this solution – they clearly did. The question was how to build it effectively.

MVP in 3 weeks? Yes, please

The team faced a challenging timeline – just three weeks to develop the MVP with a strictly limited budget. The main risk was determining the tech stack and development approach, as short-term MVP projects were new territory for everyone involved.

To save time, we opted for OpenAI API, knowing it wasn’t the cheapest solution. There was a plan to give anyone access to the tool, but there was definitely no plan to pay hundreds of dollars for the API usage.

We implemented token limitations right away – working with only the first 20 catalog items, strictly demo mode. This approach significantly reduced development time and overall MVP costs, despite the temptation to build a proprietary Enrichment Engine using local LLM libraries.

A new question emerged: why build a custom product if everyone already has access to ChatGPT? We’ll get back to this later, but here’s the spoiler alert: because our solution works better.

So, having the simple plan sketched, we began the development process. The core minimum was achieved pretty fast: backend + frontend + DB + API connection, but it turned out to be the easiest and most predictable step.

Initially, we wanted basic fraud protection through user authorization, which would also help collect newsletter subscribers and feedback. However, when testing this customer journey with internal users, we realized it could become a significant barrier to product adoption. A good insight – perhaps obvious to some from the start, but as they say, we’re all learning.

So we decided to show a preview of the results, while putting the full version behind an authorization wall. This way, users can immediately see if it’s worth spending a minute of their time or not. And it’s definitely worth it!

How we cracked the consistency problem

This is where we get to why our solution is better than just using OpenAI directly – the part you’ve probably been waiting for. While we couldn’t compete with OpenAI globally, but within our specific product? Easy win.

We quickly discovered that AI doesn’t consistently produce appropriate attributes for products. Even when hallucinations were rare, data consistency remained elusive. It simply didn’t work out of the box. Time to bring out the fine-tuning tools.

Our solution was studying what works – examining catalogs from major companies specializing exclusively in catalogs and product sales. Here’s where we should mention that datasets are everything in data enrichment projects. However much we wanted to avoid them to save time and resources, that proved impossible.

We discovered BrightData’s excellent eCommerce Datasets repository, extracted the data, and created our own set of “golden attributes” – the most important product attributes. 

amazon datasets for catalogs
Here are some examples of datasets we’ve used to train SmartCatalog

Voilà! Our product now requests specific attributes from AI and receives consistent responses.

Please don’t ask about the prompt we constructed. It’s a secret recipe locked in our vault, and honestly, it’s so lengthy you might fall asleep reading it.

Making AI with AI

We’ve been working with LLM models for programming assistance for years now, so we know both the strengths and weaknesses of this approach. This could be a whole other article, so for now let’s just say our backend developer, who had little frontend experience, quickly sketched out a sleek, functional UI/UX with AI assistance.

smartcatalog interface

It was truly impressive how fast LLM tools can elevate an experienced developer toward full-stack capabilities. The traditional boundaries between specializations are beginning to blur, with AI bridging the skills gap in real-time.

This approach significantly accelerated the development process. Instead of hiring additional specialists or spending weeks learning new skills, we leveraged AI to expand the capabilities on demand – quite fitting for a project built around AI-powered enrichment.

Big challenges for a small team

With a project team consisting of just two people – a manager and a developer – challenges emerged on both technical and organizational fronts.

The developer had to demonstrate architectural skills in a very short timeframe to avoid rework and complete everything as planned. Everyone familiar with the software development lifecycle knows about estimates, deadlines, and other realities. Things rarely go smoothly, and while most teams have future sprints for improvements and fixes, we didn’t have that luxury.

On the organizational side, the manager’s primary task was establishing a clear goal, understanding user needs, and defining the value that everyone was working toward. And, not to brag, but both team members successfully fulfilled the responsibilities despite these constraints.

This lean approach forced us to be incredibly focused. Every decision mattered, every hour counted, and there was little room for exploring alternative approaches. The time pressure actually helped concentrate efforts on what truly mattered – solving the core problem of catalog enrichment.

How SmartCatalog works

SmartCatalog operates through a simple three-step process, with little effort from the user’s side:

  1. Input: You provide existing catalog data in either XLSX or CSV format, containing product names and any existing attributes you already have.
  2. AI analysis: Once uploaded, the system quickly extracts unique product names from the text and begins enrichment. The AI engine identifies appropriate attributes for each product and ensures consistency across similar items – all within about a minute.
  3. Output: You can download the enriched catalog with ready-to-use attributes. The current beta version processes the first 20 items.

The system returns industry-standard JSON format with original data preserved, and new attributes clearly marked, making it perfect for immediate integration with websites, marketplaces, or CRM systems.

“What makes the tool special is the “wow effect” that happens when users – even in this MVP stage – see immediate insights about how they can improve their catalog. They quickly understand what elements need attention and how their data can be enhanced.

Even if someone doesn’t fully convert to using SmartCatalog as their permanent solution, the insights gained become valuable information for improving their product data management. After all, making the world better, as cliché as it might sound, is our ultimate goal.”

Stanislav, SmartCatalog’s project manager

Conclusion and future directions

What started as a solution to the “empty catalog” problem has evolved into a practical tool that helps businesses improve their product data. Building SmartCatalog in just three weeks with a team of two showed that focused effort and thoughtful use of AI can produce good results.

The golden attributes approach became our main technical advantage, giving SmartCatalog better performance than generic AI tools. By learning from existing e-commerce datasets, the system provides the consistency that raw AI responses lack.

Moving forward, we plan to:

  • Take a look into more complex LLM scheme, with on-premise basis, not to count on a third-party provider
  • Optimize the processing speed and do an overall infrastructure optimization, to be ready for a high RPS with the solid SLA
  • Produce a more specialized enrichment engine that is able to work with a certain domain at its best

Creating SmartCatalog taught us that valuable tools don’t always need to be revolutionary innovations. Sometimes, applying the right technology to solve common business problems is just as important. For our team, this is just the beginning of making product catalogs more useful for businesses.

Chapters

Related posts