In the age of digital acceleration, Latin America’s e-commerce leader, Mercado Libre, continues to set the pace. With over 492 million items sold in the first quarter of 2025 alone, the platform faces a unique challenge: how to maintain consistency, quality and speed across millions of product listings while continuing to grow. 

This is the story of how artificial intelligence became a force multiplier, facilitating smarter workflows for content generation, improving catalog consistency and empowering sellers across the region, regardless of their size or technical expertise. 

The challenge 

As Mercado Libre continued its rapid growth, the platform saw a surge in new sellers and product listings. With that scale came a familiar challenge: inconsistent descriptions, incomplete attributes, manual categorization, unprofessional images, fragmented product groupings that degrade the shopping experience. These issues not only affect user satisfaction but can also directly impact conversion rates. 

To solve this, Mercado Libre partnered with Marvik, a hands-on consulting firm specialized in building tailored AI solutions. The company approached Marvik with a clear objective:to automate and streamline the product listing process using AI in order to maintain quality standards while enabling faster and easier content creation for sellers. 

The solution: structuring and enhancing the product catalog with AI 

A comprehensive AI-driven solution was developed to automate the structuring and enhancement of product listings, delivering high quality, scalable outputs that match human curated standards. 

On the text side, natural-language models were used to identify duplicate attribute values, grouping variations like “red” and “reddish” under a single standardized label. These models also clustered semantically similar products, validated the coherence of each group, and ensured alignment between titles and listed attributes. By parsing titles and descriptions, proposed restructuring the product taxonomy—merging overlapping categories, splitting overly broad ones, and flagging missing or equivalent attributes—so the catalog evolves with the inventory. Attribute values were automatically normalized to a common unit of measure, and where details were sparse the workflow inferred likely values, generated concise descriptions, and normalized titles into a consistent pattern. All of this catalog processing ran on transformer and large language models within a generative-AI framework.. 

In parallel, computer vision models processed millions of product images to detect duplicates using perceptual hashes and embeddings, form visual clusters and identify outliers that disrupted group coherence. Within each cluster, an automated selection process chose the definitive set of product-images, prioritizing those most frequently repeated, highest quality and fully compliant with content policies. These models can also extract relevant attributes directly from visuals and flag non-compliant seller-uploaded images, while visual-similarity tools improve the accuracy of product grouping and recommendation. 

A key component in this architecture is the use of multi-modal AI agents, tools that combine insights from both text and image data to refine item clusters with higher precision. These agents also play a critical role in the moderation loop:they automatically review, approve, or reject seller-submitted suggestions for changes in product information. By enforcing catalog standards in real-time and at scale, these agents help maintain data integrity while reducing the need for manual oversight. 

Combined, these capabilities created a high-performing, self-improving catalog system capable of interpreting and elevating both textual and visual content. The result: a cleaner, more consistent and more searchable product catalog that improves discoverability, streamlines seller onboarding and refines the end-to-end user experience. 

All of this is made possible thanks to NVIDIA hardware, which provides the high-performance computer needed to train and run large-scale computer vision and NLP models in production. 

The solution: studio-quality images powered by AI 

Before After 

In e-commerce, visuals are everything. Product images shape perception, build trust and often determine whether a shopper converts or clicks away. Yet for most sellers, especially small and mid-sized businesses, creating high-quality photos is time-consuming, costly and often out of reach. That’s where Photo Studio comes in: a powerful AI solution that turns everyday photos into polished, studio-quality images without manual effort. 

Photo Studio uses a series of advanced AI models, each designed to improve product visuals. It starts by removing messy or distracting backgrounds, instantly giving products a cleaner, more professional look. Then, it smartly analyzes the image to understand what’s in it, whether it’s a face, a hand, or a pair of sunglasses, helping the system apply edits more accurately and naturally. 

One of the impressive features is its ability to recognize and isolate specific items, like a jacket or pair of shoes and refine them individually. This level of detail means garments and accessories can be enhanced with remarkable precision, bringing out colors, contours and textures that often get lost in amateur photos. 

Beyond cleaning and refining, Photo Studio also adds creative flair. It uses generative AI to automatically produce realistic, well-composed backgrounds customized to the product and its category. Whether it’s a minimal white studio setup or a lifestyle-inspired scene, the system creates visuals that elevate the product while staying consistent with brand aesthetics. And best of all, it does it autonomously, filtering out lower-quality results and ensuring every image meets a high visual standard. 

The result is a powerful blend of automation and creativity: images that are not only beautiful but also optimized for e-commerce performance. With Photo Studio, sellers can present their products with the polish of a high-end photo shoot, without ever stepping into a studio. 

With this innovation, Mercado Libre became the first marketplace to offer sellers an autonomous, AI-powered tool for improving product photography, setting a new standard for seller enablement and visual quality at scale 

The Technology 

The entire workload is GPU-accelerated to keep latency low at production scale.

• Offline processing. Data preparation and model-training jobs run on AWS g4dn.xlarge compute nodes, each fitted with a single NVIDIA T4 Tensor Core GPU (16 GB), where CUDA and cuDNN speed up intensive steps such as image analysis,text clustering and data validation.
• Online serving. The live Photo Studio pipeline runs on AWS g5.xlarge instances that carry an NVIDIA A10G Tensor Core GPU (24 GB), delivering real-time image segmentation, enhancement and generation for thousands of concurrent users without queuing delays.

To achieve this level of visual quality and responsiveness, the system relies on a robust foundation of NVIDIA technologies. CUDA and cuDNN provide the essential acceleration for deep learning tasks, ensuring that every image enhancement step from segmentation to generation runs efficiently. These tools don’t just power the engine, they elevate the experience, enabling results that are both fast and production-grade. In many ways, NVIDIA’s stack is the silent partner behind the scenes, turning AI ambition into real-world impact.

Impact and benefits 

During the rollout of Photo Studio, the platform saw a significant increase in product visits, clear evidence of how impactful high-quality content and visual consistency can be at scale. 

At the same time, the catalog structuring solution enabled smarter, more efficient content consolidation. Across tens of millions of listings, the system reduced visual fragmentation by nearly a third, not by removing products, but by intelligently grouping duplicates and variations under unified product representation. This optimization of the shopping experience led to a notable increase in engagement within affected catalog sections, reinforcing how structure, clarity, and intelligent automation can boost discoverability. 

Together, the implementation of these AI driven solutions delivered measurable improvements across multiple dimensions: 

• Catalog consistency at scale: attribute accuracy, standardized taxonomy and cleaner metadata helped optimize search and discovery. 
• Faster and easier listing creation:the new tools significantly reduced the time and effort required to publish products streamlining seller onboarding and lowering the barriers to entry for thousands of merchants. 
• Seller enablement: tens of thousands of sellers were empowered to create polished listings, regardless of their technical or creative background. 
• Improved buyer experience: clean visuals and well-structured data helped shoppers make more confident purchase decisions and find exactly the best offer to what they were looking for. 

Why this matters 

This transformation demonstrates how AI is redefining digital commerce infrastructure. At scale, consistency and speed are often in tension but with the right automation, marketplaces no longer have to compromise. 

By introducing intelligent systems that elevate both text and visual content creation, this platform created a more fluid experience for buyers and significantly lowered the barrier to entry for sellers. With structured data and high-quality visuals, product discovery improves and so does trust. 

Looking ahead 

As online commerce continues to evolve, the ability to automate complexity while maintaining brand and content standards will define the next generation of marketplace infrastructure. With intelligent AI agents embedded throughout the product listing lifecycle, platforms can operate more efficiently, serve their ecosystem more effectively, and innovate faster than ever before.