Wildberries processes over 12 million product listings and 70 million images through its filters every single day. According to Lev Nechaev, Head of Automated Moderation, the traditional "one task, one model" strategy quickly drains GPU resources and hits a performance ceiling at this scale. Even with efficient detectors like YOLO, the endless proliferation of niche neural networks for every brand becomes an infrastructure nightmare and drags out time-to-market to an absurd degree.
In business terms: instead of training a separate classifier for every minor regulatory tweak, the marketplace's engineers implemented detectors based on vector search. This allowed them to merge fragmented models into a single ensemble. The mechanics are straightforward: the system converts content into vector representations and searches for matches in a violation database. This approach covers checks for over 20 entity types—from Wibes clips to text descriptions—handling peak loads of 400+ requests per second with at least 90% accuracy.
Cutting the Hidden Costs of AI Scale
Stripping away the corporate polish, what we are seeing is a savvy optimization of unit economics. By bringing models under a common vector denominator, Wildberries can scale brand and face recognition without an exponential spike in hardware costs. This is an elegant solution to the problem of catastrophic forgetting or overfitting, where older models begin to "hallucinate" as new classes are added. Now, adapting to a new law or a major rebranding only requires updating the vector database rather than overhauling the entire architecture.
This shift proves that engineering ingenuity can protect a budget from being devoured by cloud providers.
Audit your current inference costs for specialized models; replacing a dozen narrow classifiers with one vector search system could pay for itself by next quarter. Vector databases offer more agility than retraining, allowing for near-instant updates to moderation rules. Consolidating AI infrastructure reduces the technical debt associated with managing hundreds of micro-models.