Zdeněk Neustupa

Zdeněk Neustupa

Zdeněk Neustupa is a co-founder and CTO of 24 VISION a.s., where he is responsible for the technical vision and hands-on engineering behind ANNIE, an AI-based visual quality inspection system capable of detecting product defects and configuration issues in seconds.

His work focuses on designing and deploying AI systems that operate reliably in real manufacturing environments rather than controlled laboratory conditions. As the chief architect of a company entering the product-scaling phase, he has firsthand experience with the challenges that arise when moving from proof-of-concept solutions to production-ready AI — including architectural trade-offs, hardware constraints, data drift, system integration, and long-term maintainability.

Over the past years, he has navigated the transition from a project-driven development model to a scalable product platform, balancing technological risk, customer requirements, and operational stability. This includes deploying AI at the edge, orchestrating multiple models in production, and building monitoring and retraining workflows that keep systems functional long after the initial rollout.

His primary interest lies in bridging the gap between AI research and industrial reality — ensuring that AI systems deliver measurable value, survive production constraints, and remain trustworthy over time.

When AI Meets the Factory Floor: Lessons from Real-World Visual Inspection

Presentation abstract: Training, validating, and testing an AI model on a laptop or in the cloud can feel deceptively simple. A few datasets, a promising accuracy number, a nice demo — and the problem seems solved. In manufacturing, this is usually where the real problems begin.

This talk shares hands-on experience from deploying ANNIE, an AI-based visual quality inspection system, into real production environments. We will explore what happens when AI models leave controlled lab conditions and face production realities: unstable lighting, hardware constraints, line speed requirements, changing products, imperfect data, strict uptime demands, and zero tolerance for "just restart the container."

We will discuss the practical challenges of running AI in manufacturing — from dataset drift and model versioning to deployment architectures, performance monitoring, explainability, and long-term maintainability. Particular attention is paid to the gap between model accuracy metrics and actual business value on the factory floor.

Rather than focusing on theory or isolated models, this presentation highlights the full lifecycle of industrial AI: deployment, monitoring, retraining, scalability, and integration into existing production systems. The goal is to provide realistic insights, hard-earned lessons, and concrete takeaways for anyone considering moving AI from a demo into real manufacturing operations.