Dr. Egor Kraev | AI Track | Compass AI & Data Summit 2025
summary
Claims a reliable pattern for getting LLMs to emit complex, validated objects rather than free text, using Pydantic objects as the schema, a validator-as-tool that the agent must call (agentic validation loop), and deterministic converters to SQL or chart engines. Shows working use cases: generating chart config from images, producing exact semantic layer queries for databases, creating regulated suspicious activity reports, and synthetic dataset schemas. Describes an open-source agent framework with a validator-only return path, a KV store for tools, and clear failure semantics when validation never succeeds.
AI-written summary · about
chapters (8)
- 0:00Introduction
- 3:07Structured objects for AI
- 10:13Complex validation challenges
- 15:06Validation pattern for outputs
- 18:40Ship-in-a-bottle pattern
- 24:00Reliable AI takeaways
- 26:53Q&A: Company name & use cases
- 34:44Q&A: Future outlook
related talks
12mHuman seeded Evals
35mBuilding LLM-based Agents
27mHow to Build AI Agents and Keep Your Sanity
47mBuild Trustworthy AI Agents powered by Evals
36mAutonomous Policy Validation: Building AI Agents to Analyze Logs and Identify...
29mBuilding Trust in AI Agents
30mBuilding Robust, Secure LLM and Agentic AI Workflows
01h 04mBuilding AI Applications the Pydantic Way