How Avito Integrated LLMs Into Its Moderation Pipeline

Avito processes over 20 million listings daily, with more than 95% of moderation automated. The remaining volume, amounting to hundreds of thousands of listings, is handled by several hundred human moderators.

The company integrated LLMs to improve moderation, opting for prompt-based instructions derived from internal regulations rather than RAG or full-context models. This approach balances cost, speed, and stability.

Avito uses a 35B parameter model for general tasks and a fine-tuned 8B model for complex violations. Fine-tuning is performed using LoRA on NLP layers, while vision layers remain untouched.

The system employs automated prompt optimization based on the GEPA framework, which was adapted to use token probabilities and Russian-language instructions. Final prompt candidates are validated manually.

The production pipeline includes GuardRails to prevent prompt injection and jailbreaks. If a listing is suspicious or the LLM lacks confidence, the system routes the task to human moderators.

Currently, LLMs handle approximately 70% of the listings flagged for secondary review, achieving a precision of about 94%. The remaining 30% of complex cases are reserved for future automation.