Problem
Companies pilot LLM assistants, but few reach production: the talk cites an MIT report where 5% of custom enterprise AI tools do. At Avito, employee tasks range from a quick question to a multi-step process across several knowledge bases and tools. A plain chat cannot handle those reliably, and nobody can tell which employee-built assistants are good enough to trust.
Example from the article: a corporate-center employee must find an alternative to Microsoft Office and buy software for 700 thousand rubles. She must search Confluence and chats, ask colleagues and find which 1C request to file. The route alone takes about 30 minutes.
Solution in brief
Avito built a corporate assistant, Vitalik, on its LLM portal. Users build "projects" (documents, instructions, tools), and those with a process owner, validation data and closing tools become agents. An orchestrator builds an agent tree per request, RAG and an MCP hub supply context and tools, and a four-level benchmark gates promotion.
Questions this case answers
- How to turn an internal assistant platform into production agents for multi-step processes by routing tasks among specialised agents over a shared retrieval and tool layer
- How to benchmark an internal agent before it is allowed into a shared agent system
- How to turn an internal LLM chat portal into a platform where business processes become production agents
- How to structure an orchestrator, RAG and an MCP hub for corporate agents
Tools and methods
- ADK runtime — session, state and memory layers, callbacks, MCP client
- PortalManager — in-house orchestrator that builds the agent tree
- Hybrid RAG — dense plus BM25 retrieval, reranker, nightly Confluence sync
- MCP Hub — registry of MCP servers with permissions and traces
- Benchmarks — BFCL, ruBFCL, tau2, Ragas, LLM-as-a-judge, tracked in Opik
- n8n — low-code pipelines that users build themselves
Google or a one-time email code. No password.