NVIDIA unveils multi-agent intelligent warehouse and catalogue enrichment AI blueprints

Source: NVIDIA blog post. Harnessing the Retail Catalog Enrichment NVIDIA Blueprint (left) and the Multi-Agent Intelligent Warehouse Blueprint (right).

Ageing retail systems, siloed data and rising customer expectations can now be addressed with NVIDIA's Multi-Agent Intelligent Warehouse (MAIW) and Retail Catalog Enrichment NVIDIA Blueprints.

These open-source developer references empower developers to customise AI-powered solutions for the retail value chain.

“Building with these blueprints will reduce the cost of integration and help our customers and partners enable applications fast,” said Tarik Hammadou, Director of developer relations for AI for retail and consumer packaged goods at NVIDIA.

“They unlock the efficiency and enterprise‑grade scale the retail industry needs to compete.”

The NVIDIA MAIW blueprint delivers AI support that sits above existing warehouse management systems, enterprise resource planning (ERP), robotics and Internet of Things (IoT) data, so teams gain real-time, explainable operational intelligence.

The blueprint comprises specialised agents for equipment asset operations, operations coordination, safety compliance, forecasting and document processing — all orchestrated by a central warehouse operational assistant that mirrors how warehouses actually run and turns fragmented data into proactive decision-making.

An ongoing issue within warehouses, for example, is a disconnect between the IT and operational technology (OT) layers. This gap prevents managers from easily handling problems such as accurately measuring product inventory, efficiently pinpointing technology issues and deploying enough workers to areas that need extra help.

“The idea of having an agentic AI layer on the IT or OT level is not efficient, but having agents in between IT and OT allows the AI agents to act as the coordinators,” said Hammadou.

With the blueprint, a supervisor can ask “Why is packing slow?”, triggering the assistant to analyse equipment status, tasks queues and staffing data to identify bottlenecks with supporting evidence and recommendations.

The blueprint also provides production-grade capabilities — including role-based access control and guardrails to keep recommendations within policy — so operations teams can trust AI to help coordinate real equipment and safety-critical decisions.

By targeting metrics to detect and resolve issues and safety incidents, as well as ensure on-time order fulfillment and service level agreement adherence — MAIW helps warehouses move from constant fire drills to more predictable, data-driven shifts.

The Retail Catalog Enrichment NVIDIA Blueprint, on the other hand, can help businesses achieve richer, more accurate product onboarding, as well as localised marketing.

Retailers often face a “sparse data” problem with product images: minimal or inconsistent text associated with the image. Teams can spend a lot of time writing titles, descriptions and attributes for each product, then customising them for each market and campaign.

The blueprint addresses this by using generative AI to create high-quality, structured, localised and brand-aligned product content at scale.

Take a retailer trying to update their online storefront with a set of mugs. With an NVIDIA Nemotron vision language model (VLM), part of the Retail Catalog Enrichment Blueprint, the photos of the mugs are fed through the VLM to generate product metadata such as colour, material, capacity, style and use cases.

From single images, the system can then produce localised product titles and descriptions, extract and normalise attributes for search and recommendation systems for improved search engine optimisation (SEO) generative engine optimisation (GEO), and also create culturally-relevant 2D lifestyle imagery as well as interactive 3D assets. Behind the scenes, an AI “judge” checks outputs for quality and consistency.

Brand voice, tone and taxonomy instructions can be layed on via prompts, alongside the product image and a target locale. The blueprint uses brand guidelines to generate enriched product titles and descriptions, localised categories and tags, and culturally-appropriate lifestyle image variations tailored to that intent.

Companies are already creating their own products with the help of NVIDIA’s retail blueprints. Global tech consulting firm Grid Dynamics, which has a presence in India, has built a catalogue enrichment and management system that increases the accuracy of item content and status of stock keeping units (SKUs) for large retailers, using the Retail Catalog Enrichment NVIDIA Blueprint. 

The challenge for bigger retailers with massive product catalogues is that attributes can be missing or incorrect. Onboarding new vendors with differing catalogue structures can further jumble the data — leading to inaccurate sales, frustration and, eventually, a loss of customer loyalty. 

“The quality of the search and the quality of the browsing experience for customers directly depends on the quality of the catalogue data,” said Ilya Katsov, CTO, Grid Dynamics. 

“It’s a very critical problem for all retailers with a digital presence to ensure their catalogues have as rich and consistent of attributes as possible — and our solution automates this so they don’t need to do manual reviews.”

This is where Grid Dynamics’ solution comes into play.

“Our solution makes product catalogues more discoverable while giving brands the ability to enforce their business rules at scale,” said Dan Guja, Principal software engineer at Grid Dynamics. 

“With AI-driven business rules applied across the catalogue, brands can improve data quality, sharpen customer intent signals and surface products customers actually want.” 

“The next step is embedding a physical AI layer into warehouse and store operations, enabling intelligent agents to see, reason, and act on real-world inventory and supply-chain challenges,” said Hammadou. 

“By training physical agents with capabilities like computer vision, we’re moving toward more adaptive and autonomous operations.”