AI Technologies
We create our own ML technologies, integrate them into T‑Bank products, and share them with the market
Natural Language Processing
We teach machines to understand natural speech and chat with users thoughtfully and intelligently
Our technologies
We utilize the full spectrum of NLP technologies, from regexp to our own LLM, in a variety of ways:
- Creating voice and text robots, a powerful search engine, and smart tools for internal operations
- Processing communication on social networks
- Preprocessing texts for further use in personalization, risk modeling tasks, and much more
Speech analytics
We create tools to automatically control the quality of customer service. We cluster, classify, and highlight problematic parts of dialogues
Search systems
We’re developing a semantic search feature, building embeddings, breaking down search queries into semantic blocks, and solving search tasks using knowledge bases with billions of documents
Dialog systems
We use NLP technologies to solve users’ problems by maintaining the quality of customer service and shortening the time it takes to resolve their cases
GEN‑T: A Family of Custom Large Language Models
We’re not making a GPT analog. Rather, we create the best solutions for each specific domain
Efficient inference
High tokenization density and expanded model context capacity make it possible to solve a wider range of tasks
Cutting‑edge alignment
We integrate the latest tuning methods into the product, creating safe LLMs with predefined properties
Agents and function calling
The ability to use tools and dynamically decompose tasks provides a large degree of flexibility in solving them
Adaptation and pretraining in Russian
We manage the facts the model knows, which helps reduce hallucinations and ensures the relevance of the model’s understanding in the respective domain
Voice Technologies
We develop a basis for analytics and the creation of a human‑technology interaction interface in the voice domain using precise, fast, and secure models
Voice conversion
We use generative adversarial networks to modify the timbre of a speaker’s voice while leaving the linguistic and intonational information unchanged
Speech synthesis
We strive to generate expressive and natural speech that is indistinguishable from that of an actual person
Speech recognition
We develop solutions that accurately recognize words and subtleties in a person’s voice to determine a person’s age, emotions, and intonation. Our systems are adapted to telephony and various accents
Speaker recognition
We’re able to verify a person’s identity with high accuracy by their voice, extracting and analyzing their unique characteristics
Anti‑fraud Technologies
We develop solutions that protect our clients and T‑Bank’s IT infrastructure
Guarding Against Risks
We leverage the full spectrum of big data analysis and machine learning, from time‑tested methods to state‑of‑the‑art deep learning technologies
Our highload ML microservices constantly analyze data streams to instantly detect and prevent any unsanctioned attempts at accessing our clients’ money or attacks on our IT infrastructure
Neuroshield
We detect and stop fraud right during the call
Safe, fraud‑free conversations
Our ML systems carefully analyze the sound waves of a person’s voice, paying attention to hundreds of technical details that could indicate fraudulent activity
Instant fraud protection
As soon as the algorithm identifies any sign of fraud, the call is immediately ended, preventing scammers from continuing the call and completing their scam
Always a step ahead of scammers
Scammers are always looking for new ways to commit fraud, but our system isn’t standing idly by — it learns from every call, always staying one step ahead to make sure our clients are safe
Computer Vision
We create image processing services that save time on routine tasks
Biometric face recognition
This guarantees an additional layer of security for our users’ transactions
TFusion
We use and fine‑tune generative models to accelerate designers’ workflows and personalize visual communication with our users
OCR and entity extraction in documents
We build OCR systems for a variety of document types: multi‑column texts, tables, and distorted documents
We process millions of documents per year, streamlining document routing
We extract entities, such as names, dates, and numbers, using both recognized text and image‑based data
Image search
We’ve created a search feature that gives users the ability to quickly and precisely find products that are similar to images that they provide
We determine users’ intent in the chat using images they submit, improving the «intelligence» of the chat, how long it takes to resolve their problem, and the degree of automation
Recommender Systems
We develop technologies that personalize the T‑Bank ecosystem for each user, improving their client experience
Algorithmic feeds
- Development of recommender systems that personalize news and content feeds, social media posts and social subscription graphs
- Building ML algorithms for ranking and blending recommendations from diverse content sources into a unified feed
- Implementation of approaches to content and user embeddings to address the cold‑start problem
- Introducing RL and multi‑armed bandits for recommender systems trained on long‑term feed engagement metrics
Search personalization
- Search result ranking of the same query personalized for each client
- Embeddings for semantic search based on queries and result documents
User recommendations
- Item2item recommendations: product recommendations, cart‑based upsell suggestions
- User2item recommendations: personalized recommendations for users regarding securities, stores, products, content, etc.
- Next‑item and next‑basket recommendations: for predicting the next target action or generating various recommendations. Tasks include next‑basket prediction and recommended collections
Time Series
We provide businesses with precise forecasts and help them find insights
Auto ML
It optimizes the process of working with time series by proposing effective initial settings based on past experience. This reduces the time for model selection and testing, and provides quality forecasts without the involvement of specialists
TimeseriesDB and edge computing
They allow data analysis to be conducted directly on the device, speeding up results while minimizing delays associated with data transmission. This makes analytics more effective for real‑time apps
Pre‑trained models
They eliminate the need for lengthy and costly training, quickly delivering high‑quality results without significant investments in computational resources