T-Brain: A Large Commerce Model Built for Every Commerce Decision

‍T-Brain is Topsort’s enterprise-grade AI platform for commerce, built around a Large Commerce Model that learns from products, shoppers, context and business outcomes. Pretrained across Topsort’s commerce network and fine-tuned to each business, it brings one shared intelligence foundation to search, recommendations, advertising, personalization and agents.

Commerce is changing how products are discovered.

Traditional commerce systems were built around keywords, product grids, and fixed placements. Search, recommendations, advertising, and merchandising often evolved as separate systems, each with its own models and view of the shopper.

That architecture becomes less effective as discovery grows more contextual, ads and organic experiences converge, and AI agents begin participating in shopping.

Large language models can understand what a shopper says. Commerce AI also needs to understand what shoppers do, how products relate, what is available, how price affects demand, and which decisions drive business outcomes.

That is the problem T-Brain is built to solve.

A Large Commerce Model that understands more than language

T-Brain is Topsort’s AI platform for commerce, built around its foundational Large Commerce Model (LCM).

Unlike a general-purpose language model, a Large Commerce Model learns from the signals behind commerce decisions.

A click is a signal. So is a product a shopper chooses not to click. Searches, purchases, product relationships, inventory, pricing, advertising performance, and marketplace economics all provide information about what shoppers want and what a business should do next.

Context changes those signals. The same shopper can behave differently depending on the query, page, time, device, location, and products being shown.

T-Brain learns from aggregate, anonymized patterns across Topsort’s commerce network, combining impressions, clicks, purchases, catalog information, and interaction context. From those signals, it can learn patterns around shopper intent, product quality, price sensitivity, and behavior.

The model is then fine-tuned to each business’s own catalog, shoppers, and commerce data.

Pretrained across commerce, adapted to each business

Pretraining and fine-tuning give T-Brain both broad commerce knowledge and business-specific intelligence.

A model trained only on one marketplace is limited to the data that marketplace has generated. T-Brain starts with patterns learned across more than 100 marketplaces, providing a stronger foundation before adapting to an individual business.

Teams can use product and interaction data already sent to Topsort and add organic events, catalog attributes, categories, and other first-party signals.

The base model learns from aggregate, anonymized patterns rather than transferring one retailer’s proprietary catalog or data to another. Once fine-tuned, the resulting model is specific to that business and its use cases.

In Topsort’s evaluations, this approach consistently outperformed models trained on individual marketplaces alone.

One commerce AI model, four core tasks

T-Brain makes the same commerce intelligence available through four core tasks:

Retrieval selects the products most relevant to a shopper and context. The assortment can change based on the query, page, channel, user, or moment.

Prediction estimates outcomes such as product quality, click probability, or purchase probability.

Ranking orders organic products, sponsored products, or both according to the business outcome a team wants to optimize.

Understanding provides product and shopper embeddings that capture relationships learned from real commerce behavior and can be used in other models and systems.

Together, these capabilities create a shared AI foundation for search, recommendations, advertising, personalization, merchandising, and agentic commerce. Instead of building a separate model for every surface, teams can apply the same underlying understanding of shoppers and products across the commerce experience.

From commerce data to a deployed AI model

T-Brain brings the model workflow into one platform, from data and fine-tuning through evaluation, deployment, and monitoring.

Teams provide catalog, category, and user-event data, select the task they want to solve, and fine-tune the foundational model on their own information.

Before deployment, they can evaluate the adapted model against their existing baseline. Once deployed, T-Brain provides an endpoint and ready-to-use API code that can be connected directly to a commerce experience.

Teams can then monitor request volume, latency, errors, and model performance.

Much of the computational work happens before a shopper makes a live request. Historical behavior, catalog changes, and event streams contribute to the model ahead of time, allowing predictions to be served in under 50 milliseconds at P95.

T-Brain results in production

T-Brain builds on models Topsort has already used to power auction decisions such as product quality, conversion, and order value. Those capabilities are now available for broader commerce applications.

In production tests:

• Prediction models increased conversion rate by 26% versus the existing baseline.
• Retrieval models increased click-through rate by 21% versus baseline.

Separate evaluations compared T-Brain models that were pretrained and then fine-tuned with models trained only on an individual marketplace.

For prediction, fine-tuned models improved click prediction across all seven marketplaces tested, averaging +5.4 AUC points. Purchase-after-click prediction improved by an average of +10.6 AUC points, with individual gains exceeding 15 points.

Retrieval models also improved ranking quality across all eight catalogs tested, ranging from 130 to 894,000 products. Smaller marketplaces saw gains of up to 55%, while the largest catalog saw an 80% improvement in ranking quality.

On the smallest catalog, the correct product ranked first 62% of the time, up from 46% for the marketplace-specific model.

Production results measure changes in shopper behavior, while these evaluations measure model quality. Both show the value of combining broad commerce learning with knowledge specific to each business.

One intelligence layer across commerce

Commerce stacks often treat search, recommendations, advertising, and personalization as separate problems. Each system learns independently and operates from a partial view of the shopper and the business.

T-Brain provides a shared commerce intelligence layer across them.

Search results can adapt to a shopper and query. Home, category, and product pages can personalize around context. Advertising can draw on the same signals as organic discovery. Communications can select more relevant products. AI agents can interact with a catalog using intelligence built specifically for commerce.

Instead of maintaining separate models for each decision, businesses can build on one AI foundation that understands how those decisions connect.

That is the idea behind T-Brain — and the foundation of Topsort’s expansion into Commerce AI.

Explore more here or book a demo.

‍

Frequently asked questions

What is T-Brain?

T-Brain is Topsort’s commerce intelligence platform, built around a foundational Large Commerce Model designed specifically for commerce. It learns from shopper behavior, product data, context, and business outcomes, then makes that intelligence available across retrieval, prediction, ranking, and understanding.

How is T-Brain different from a general-purpose LLM?

LLMs are primarily trained to understand and generate language. T-Brain is designed to understand the signals behind commerce decisions, including impressions, clicks, purchases, product relationships, inventory, pricing, shopper context, and marketplace outcomes.

That allows it to answer a different set of questions: which products are most relevant, how likely a shopper is to click or buy, and how products should be ranked for a particular objective.

How does T-Brain adapt to each business?

T-Brain starts with patterns learned from aggregate, anonymized commerce data across more than 100 marketplaces. Businesses can then fine-tune the model using their own catalog, shopper interactions, organic events, and other commerce data.

The result combines broader commerce learning with the signals and behavior specific to an individual marketplace.

What can teams use T-Brain for?

T-Brain supports four core tasks:

Retrieval to identify the most relevant products for a shopper or context.
Prediction to estimate outcomes such as clicks, purchases, or product quality.
Ranking to order organic products, sponsored products, or both around a chosen objective.
Understanding to generate product and shopper embeddings that can also be used in a business’s own systems.

These capabilities can support search, recommendations, advertising, personalization, merchandising, communications, and AI-agent experiences.

Does T-Brain share one retailer’s data with another?

No. The foundational model learns from aggregate, anonymized commerce patterns. Topsort does not carry one retailer’s specific catalog or proprietary data into another retailer’s model.

When a business fine-tunes T-Brain on its own data, the resulting model is specific to that business and its use cases.

Is T-Brain built for real-time commerce?

Yes. The computationally intensive work happens ahead of the live request, while the deployed models are optimized for real-time use. At launch, Topsort reported predictions served in under 50 milliseconds at P95.

T-Brain is already being tested in production, where prediction models delivered +26% conversion rate and retrieval models delivered +21% click-through rate versus their respective baselines.