LLM Translation: Building Real-Time Multilingual Features with Claude

adamfallon.com Posts

Using LLMs as a Translator November 19, 2024

Introduction

Travel disruptions are already stressful. If the message is in the wrong language, it is worse. At Trainline we saw a clear gap: disruption messages were not translated for customers traveling internationally.

We asked a simple question: can LLMs close that gap quickly enough for real-time use?

97cbf6a5f796712499dd890006868493ab636ace.jpg

Fixing this would improve customer retention and give us a reusable foundation for future multilingual features.

The Hypothesis

Our AI Lab strategy was to build small, useful services that could later plug into bigger systems. Real-time translation fit that model.

We hypothesized it would:

  • Differentiate Trainline from competitors like Omio and Uber
  • Reduce disruption-related frustration and improve retention
  • Create a translation capability we could reuse across products

We also wanted evidence that LLMs could handle low-resource translation. The paper From LLM to NMT: Advancing Low-Resource Machine Translation with Claude provided strong results for Claude 3.5.

56d8b8ed481326f5b94c697220c986bb7c58b709.png

The Approach

We focused on three areas:

  • Prompt design
    • Inject Trainline-specific vocabulary for context
    • Adjust to source and target languages dynamically

6f1c051a775165293be7019bd642b499edd59d4c.png

  • Technical implementation
    • Embed translations directly in disruption notifications
    • Keep the system language-agnostic for future expansion
  • Evaluation
    • Offline: BLEU scores for translation quality
    • Online: A/B tests for satisfaction and retention

d2c4caadf185c34e8d4c05bfede876f8b3efa8a0.png

6b7d12e5a4a88443f7a58d3b6514f38e5f414f84.png

What we built - A C# Library called Trainline.TranslationService

  • A database of Trainline-specific terms to inject into the prompt
  • A Python backend service that connects to AWS Bedrock (Claude 3.5)

a5c134a1db65252ba5bef8c8453a13af59f2630e.jpg

27df9b79355639e6de383f3f7e2e94dd10141501.png

e4a2ec38f0502c8ca28b8dedb8bd7ce50aba6466.png

Outcomes

  • Fewer complaints about unreadable disruption messages
  • More positive feedback on clarity during disruptions
  • A reusable translation service we can plug into new products

Learnings

Real-time translation is hard: accuracy, latency, and cost all matter. But careful prompt design plus strong LLMs made it viable. The bigger lesson was product focus: solving a real pain point creates momentum for future features.

7f0cdcdd878901e2d555823200fcbd4a09c12e42.png

Recap

By fixing untranslated disruption messages, we improved the immediate customer experience and built a foundation for future multilingual features. Real-time LLM translation is more than a nice-to-have; it is a step toward a global travel assistant.

If you're interested in building production systems, check out my post on how I built this blog with Emacs and Docker.