Perseus от Т-Банка: фреймворк ML-персонализации по всей экосистеме сервисов

T-Bank unveiled Perseus on July 23, 2026 — a universal personalization framework that combines all user actions across the ecosystem's different services into a single sequence of events and makes it possible to boost business metrics by 3% to 17%.

What problem does Perseus solve

Perseus closes the gap between the data of dozens of T-Bank services — from supermarket purchases to mortgage applications. Previously, each ML model was trained only on the data of the service where it was used, so it "saw" the user in fragments rather than as a whole.

The problem was compounded by data heterogeneity: teams collect events independently, each service has its own schema, and new products have no behavioral history at all — a classic "cold start." According to T-Bank's description, this is precisely what prevented personalization from being transferred between ecosystem products and forced each team to build its own model from scratch.

• Author — Oleg Lashinin, Head of Recommender Systems at T-Bank
• Problem: each service has its own event schema and its own isolated model
• New services have no user history — a "cold start"
• Solution: a single sequence of all user actions
• Business metrics growth — from 3% to 17% across different tasks

How the unified event stream works

Perseus glues heterogeneous events from all services into a single chronological sequence and builds a complete picture of the user on that basis. According to the developers, this approach makes the framework a flexible tool for working with ecosystem-wide data and makes it possible to scale ML personalization across the company's different products, rather than retraining a separate model for each service.

It is precisely the unified sequence that solves two tasks at once: it brings events with different schemas to a common format and transfers behavioral signal from mature services to new ones. As a result, the model gets context about the user even before they perform their first action in a specific product.

"When there are dozens of services in the ecosystem, you want the model to understand the user as a whole, not in fragments," —

Oleg Lashinin, Head of Recommender Systems at T-Bank, in the company's blog on Habr.

What this gives the business

The main result is metric growth even where personalization previously didn't work. According to T-Bank, business metrics grew from 3% to 17% across different tasks, and the effect persisted even for users who hadn't made a single purchase in a specific service.

This is a key difference from classic recommender models: without behavioral history, they're helpless. Perseus, on the other hand, relies on user actions in other ecosystem services and delivers personalization from the very first touch with a new product. For a company with dozens of services, this means that launching personalization in a new product no longer hinges on accumulating data.

What this means

Large fintech ecosystems are moving toward a "unified behavior profile": instead of dozens of isolated models, there's a single event stream that all services learn from. For T-Bank, this is a way to quickly enable personalization in new products that don't yet have accumulated history, and to squeeze out up to 17% growth in business metrics.

Frequently asked questions

What is Perseus from T-Bank?

Perseus is a universal personalization framework that combines user actions across all services of the T-Bank ecosystem into a single sequence of events. ML personalization for the company's different products at once is built on this shared data.

How much does Perseus improve metrics?

According to T-Bank, business metrics grew from 3% to 17% across different tasks — including for users without a single purchase, where standard models don't work due to the lack of history.