About event

A complex search query, a pile of structured requirements, and a system built on one or two sequential LLM calls. he prompt grows to several screens, costs increase, and the output varies every time. Based on a real-world production case in the fintech domain, Vladyslav Bocharov from SPD Technology will show how to break that monolith into several agents with specialized skills, and what it means for reliability and accuracy in production.

What we’ll cover:

• why sequential LLM calls break down on complex search queries;
• how to split a complex task into skills so that each one has clear boundaries and the right level of granularity;
• how to assess reliability when sub-agents depend on each other and their errors are correlated;
• how to bring agent outputs together: passing context, resolving contradictions, validating results;
• monolith vs. multi-agent architecture: how they compare on quality, latency, and operational costs;
• practical guidance: when it makes sense to split a task into separate skills, and when it doesn’t.
  
  
  For whom:
  
  AI/ML and MLOps engineers, data architects, tech leads, and engineering managers working with GenAI, LLM systems, and agents.