Mayank Soni is a senior AI engineer at Trainline who builds customer-facing agentic systems and develops practical safeguards for deploying them. His work spans conversational-AI research, language-model evaluation, and the operational challenges of delivering reliable travel assistance.
Earlier, at Trinity College Dublin’s ADAPT Centre, Soni researched how dialogue systems could make conversations more engaging and coherent. In 2021, he co-authored a study of self-disclosure in neural dialogue that reranked candidate responses during inference. His subsequent research on topic transitions analyzed 215 annotated conversations, and he released the associated conversation-annotation dataset and code.
In 2023, Soni and Vincent Wade compared ChatGPT-generated and human-written summaries using automated metrics, blinded human assessments, and classifiers. Human reviewers struggled to distinguish the two, while automated classifiers could separate them—a finding that underscores the difficulty of judging generated text by apparent fluency alone.
At AI Engineer Europe 2026, Soni described how Trainline manages the cost and reliability of production AI:
- Model switching with measurable safeguards: Evaluate cheaper or more token-efficient models against existing systems before deployment, then verify their performance against live interactions.
- Continuous offline and online evaluation: Test new customer-facing features before release and monitor production behavior afterward to catch regressions.
- Tool-level observability: Inspect individual tool calls and token consumption to diagnose complex agent workflows before and after deployment.
Soni applies these practices to Trainline’s travel-assistance environment, where agents must operate reliably amid complicated journeys, ticket rules, changing models, and real customer expectations.
