What is it about?
When you book a train ticket online and pick your seat from a diagram, someone had to build that diagram first. On Ukraine's national railway platform, which serves over 20 million passengers a year and covers more than 150 carriage types, creating one of these seat maps by hand took an engineer about 24 hours, and the change could only go live with the next software release. This study built a system where two AI components share the work. The first reads the official carriage drawing and reasons about what it shows; the second turns that reading into the exact configuration file the booking system needs. A separate checking step then verifies the result against strict rules before anything reaches passengers. Across 20 real carriage drawings, the process dropped from roughly 24 hours to 15 minutes.
Featured Image
Photo by Dasha Bakanova on Unsplash
Why is it important?
Most discussion of AI in production systems stops at the point where the model produces an answer. The harder question is whether that answer can be trusted to reach real users. This work addresses that directly: the AI never publishes anything on its own, because every generated configuration must pass a deterministic validation step that checks structure, format, and correctness before deployment. Across all test cases, no errors caused by model hallucination reached production. The practical effect is that non-technical staff can now update carriage layouts themselves, without waiting for an engineer or a software release. The supporting infrastructure runs serverless for under $5 a month, and the core booking application was never modified. The same pattern — separating the AI's reasoning from its execution, then gating the output behind formal validation — applies well beyond railways. Any system that represents a physical object as a digital interface schema faces the same problem: airline cabins, stadium seating, cinema halls, parking structures.
Perspectives
I started with the obvious design: one model reads the carriage drawing and writes the configuration. It mostly worked, and "mostly" is not good enough when the output decides which seats 20 million passengers a year can book. So the architecture became two layers — reasoning separated from execution, so failures are legible — with deterministic, rule-based validation as the gate. I deliberately avoided using a second model as the reviewer; that trades one probabilistic output for a probabilistic opinion about it. The result I did not expect: the biggest win was not the 24 hours to 15 minutes. It was that updating a layout no longer needs an engineer or a release cycle.
Ivan Dobrovolskyi
Western Governors University
Read the Original
This page is a summary of: Development of a two-layer agentic architecture for automated generation and validation of wagon user interface schemas using multimodal large language models, Eastern-European Journal of Enterprise Technologies, June 2026, Private Company Technology Center,
DOI: 10.15587/1729-4061.2026.363741.
You can read the full text:
Contributors
The following have contributed to this page
