VK AI Team Reports on the Development of the Discovery Platform

The VK AI team reported on the development of the Discovery platform on Habr (29.10.2026) — a unified layer for recommendations, search, and advertising across all VK products: 15 new recommendation verticals launched in half a year, a new recommender deployed in a day, and 70,000 CPU cores freed up through end-to-end runtime optimization.

Compiled by Sergey KostenchukPublished 2026-10-29Updated 2026-10-29

2026-10-29 Coding

💻 VK AI Team Reports on the Development of the Discovery Platform

A report on the unified layer for recommendations, search, and advertising for VK products was published on Habr (29.10.2026). In half a year, 15 new verticals were launched, a new recommender starts in a day, and about 90% of ranking tasks are solved via configuration without code. The Discovery Portal provides automatic rollback when metrics degrade, transformer GPU inference achieves millisecond latencies, and runtime optimization freed up 70,000 CPU cores. Impact: +20% watch time on the VK Video main page, +11% in the VKontakte feed, and +35% track additions in VK Mix.

🌍 Configuration over code and GPU inference are becoming the standard for recsys, while optimizing the shared runtime saves resources for all VK products at once, not just a single service.

👤 The architecture of industrial recsys is visible: dynamic batching of GPU requests, latency control by percentiles, and degradation instead of failure when a model fails.

Source 1: https://habr.com/ru/companies/vk/articles/1101174/

Sources