Back office ops · Production

Zalando uses LLMs to migrate UI component libraries across 15 B2B applications

The problem

Zalando's Partner Tech department had two distinct in-house UI component libraries causing inconsistent partner experience, duplicated design and development efforts, and increased maintenance complexity across 15 B2B applications.

First attempt

Initial LLM migration attempts without structured prompting produced inconsistent results; interface-only prompting yielded low accuracy; and automated mapping generation introduced flaws such as incorrect visual size equivalencies between the two libraries.

Workflow diagram · grounded in source
1
Component prompt crafting
trigger
“we crafted the transformation prompts for migration by providing the source code of our components to the LLM. These initial instructions included component interfaces, transformation rules, and example migrations. We utilized continue.d…”
2
Manual prompt verification
human_review
“we included manual verification of the prompts, for example, fixing the size mapping if they were not accurate”
3
LLM file transformation
ai_action
“We developed a Python based migration tool using the llm library's conversation API. The tool processed each file in the given source directories and applied LLM-powered migrations for the components present in the file”
4
Large file continuation
ai_action
“utilizing the conversation API and passing "continue" as a prompt whenever the content was cut off. This allowed the LLM to pick up where it left off and complete the transformation”
5
Automated regression testing
validation
“we implement automated testing using LLM-generated examples. These examples served as validation tools and regression tests, helping us catch unexpected changes during the migration process”
6
Human code and visual review
human_review
“code reviews and thorough visual testing would be needed for catching subtle issues that LLMs might introduce”
Reported outcome

The LLM-powered migration achieved over 90% accuracy across large file volumes and cost less than $40 per code repository, delivering significant time and resource savings.

Reported metrics
Component migration accuracymore than 90%
Migration cost per repositoryless than $40
Processing time per file30 and 200 seconds per file
Time and resource savingssignificant time and resource savings
Show all 6 reported metrics
component migration accuracymore than 90%
migration cost per repositoryless than $40
processing time per file30 and 200 seconds per file
time and resource savingssignificant time and resource savings
manual fixes after LLM migrationreduced the amount of manual fixes needed
applications in scope15
Reported stack
GPT-4ocontinue.devllmPython
Source
https://engineering.zalando.com/posts/2025/02/llm-migration-ui-component-libraries.html
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

The LLM-powered migration achieved over 90% accuracy across large file volumes and cost less than $40 per code repository, delivering significant time and resource savings.

What tools did this team use?

GPT-4o, continue.dev, llm, Python.

What results were reported?

Component migration accuracy: more than 90%; Migration cost per repository: less than $40; Processing time per file: 30 and 200 seconds per file; Time and resource savings: significant time and resource savings (source-reported, not independently verified).

What failed first in this deployment?

Initial LLM migration attempts without structured prompting produced inconsistent results; interface-only prompting yielded low accuracy; and automated mapping generation introduced flaws such as incorrect visual size…

How is this back office ops AI workflow structured?

Component prompt crafting → Manual prompt verification → LLM file transformation → Large file continuation → Automated regression testing → Human code and visual review.

Related back office ops cases