Just Now Possible Building Trainline’s AI Travel Assistant: How a 25-Year-Old Company Went Agentic
16 snips
Oct 30, 2025 Join David Eason, Principal Product Manager at Trainline; Billie Bradley, Product Manager specializing in AI; and Matt Farrelly, Head of Machine Learning Engineering. They dive into the development of Trainline’s AI Travel Assistant and the unique challenges they faced. Discover how combining AI with deep industry knowledge enhances user experience and the importance of well-structured guardrails. They also discuss innovative evaluation methods and the potential of scalable, real-time support for travelers, even in a traditional company.
AI Snips
Chapters
Transcript
Episode notes
Mix Static Content With Real-Time Data
- Trainline combined a large curated corpus with real-time APIs so the assistant can answer static and live questions.
- Real-time position and incident data plus curated content makes context-aware responses possible.
Layer Guardrails Around LLM Outputs
- Do implement guardrails both before and after generation: contextual grounding checks, toxicity checks, and tool validation.
- Use a specialized LLM to judge retrieved documents' relevance before feeding them to the orchestrator.
Prevent Infinite Agent Loops
- Limit reasoning loops and break monolithic tool inputs into smaller tools to reduce recursion and hallucination.
- Put hard iteration caps in the orchestrator so agents stop after a fixed number of tool calls.