“Reliable LLM systems for Real-World Data Extraction”
Large Language Models show impressive results in demos, but extracting structured data from real-world inputs such as emails, documents, and semi-structured text remains a major challenge in production.
This talk focuses on building reliable, production-grade LLM systems for data extraction. I will discuss why prompt-only approaches often fail and present practical architectural patterns that improve robustness: preprocessing, multi-step extraction, schema validation, error handling, and evaluation strategies.
The session is based on real production experience and highlights engineering trade-offs rather than experimental prototypes or hype-driven examples.
Bio
Andrii is a Senior AI Engineer at DataArt, working on production-grade AI systems, including LLM-based data extraction, RAG pipelines, and AI-powered automation.
His work focuses on building reliable AI systems beyond demos — dealing with noisy real-world data, evaluation challenges, and system robustness in production environments.