Wink and Rostelecom: Entertainment Agent (AI Assistant)

The main articles are:

• Speech Recognition (Technology, Market)
• Speech technology: On the path from recognition to understanding
• Data mining Data mining

2025: "Entertainment Agent" Beta Launch

Wink and Rostelecom are launching an open beta test of the Entertainment Agent service, which is integrated into the platform and application interface. This service based on artificial intelligence technologies is designed to solve the main problem of the viewer - to quickly find a movie or series in the mood and situation from the vast ocean of content. Analytics proves that long searches demotivate a person, cause frustration and eventually reduce the number of views. Rostelecom announced this on December 17, 2025.

According to research, the longer a viewer searches for the right content, the less likely they are to move on to real viewing. Every minute matters. For example, when the content is found within two minutes, about 60% of the audience starts viewing, and if the search is delayed to four minutes, then the share decreases to 20%. If more than eight minutes are spent on the search, then the views are steadily striving for 0%. Therefore, all online cinemas seek to develop recommendation systems in such a way as to leave the viewer on the platform and lead to real viewing of the presented content.

Шаблон:Quote 'author=said Anton Volodkin, CEO of Wink online cinema (a joint venture between Rostelecom and NMG).

The service was developed by the combined team of the Rostelecom artificial intelligence competence center and the Wink online cinema. By the beginning of work in open mode, "Agent" underwent in-depth multi-month training. Each Wink content item received special digital tags to speed up the search for the desired parameters. Neural networks were taught to look for answers to arrays of relevant information, including movie sites with data on films and TV shows, reviews and reviews of viewers (including emotional assessments to select the series for the mood), posts on social networks and messenger channels.

Шаблон:Quote 'author=noted Dariy Khalitov, Deputy President - Chairman of the Board of Rostelecom.

The AI agent makes the process of finding entertainment content much easier and faster, as he understands fuzzy requests ("a movie to watch in the evening" or "a series for romantic viewing"). An important prerequisite for the development of the service was the dynamic development of LLM models (large language models) that can work with natural languages, and in the case of the Entertainment Agent, the current features of Russian speech are taken into account. For the high-quality operation of the service, a combination of neural networks is used, several classifiers have been trained, and LoRA technology is used for different types of scenarios (Low-Rank Adaptation - configures the answers of pre-trained neural networks for specific tasks). For fast and high-quality semantic (semantic) search for answers, a modification of the Reindexer database was carried out - a domestic open source solution. "Entertainment Agent" supports the context of the conversation - you do not need to re-enter the data that the user reported earlier.

Preliminary testing of the service in focus groups received high marks: 84% of participants considered the product unique to the market, and the total percentage of the value of the AI agent reached 81% (according to the CSI/UMUX user experience metrics). Entertainment Agent training will go on continuously, now including at the request of large groups of real Wink users.

The Entertainment Agent is 100% powered by Rostelecom's technologies and IT infrastructure, including Neurosluse (the corporate hub of leading neural networks) and the computing power of the Turbo Cloud platform (high-performance GPU processors necessary for quick processing of requests by neural networks).

During beta testing, "Entertainment Agent" will be available authorized to adult users mobile applications in Wink for and. operating systems iOS Android In December 2025, it is planned to involve 20 thousand users in testing.

__________________________________________________