Building with LLMs: Real-World Projects & Challenges Faced

In this deeply practical session from the Artificial Unintelligence Conference 2025, Vladyslav Fliahin shares what it actually takes to build and deploy LLM-powered systems in production. Moving beyond model-centric thinking, he breaks down the real challenges teams face — from messy inputs and retrieval failures to shifting requirements, user expectations, and the constant need for iteration.

Vladyslav shows why the success of LLM applications depends less on model size and more on the surrounding ecosystem: context pipelines, orchestration logic, guardrails, evaluation loops, and domain-specific constraints. Through real examples from shipped projects, he explains how teams can navigate the unpredictability of LLM behavior and design systems that remain reliable as they evolve.

Key themes include:

Why most failures come from edges, not model weaknesses

The role of continuous iteration in real-world LLM deployments

How retrieval, data, and context assembly shape output quality

Designing domain-specific guardrails and validation logic

Aligning LLM behavior with human expectations and workflows
Vladyslav shows why the success of LLM applications depends less on model size and more on the surrounding ecosystem: context pipelines, orchestration logic, guardrails, evaluation loops, and domain-specific constraints. Through real examples from shipped projects, he explains how teams can navigate the unpredictability of LLM behavior and design systems that remain reliable as they evolve.