← back

Shipping complex AI applications — Braintrust & Trainline

4.5K views · May 01, 2026 · 98:34 min · Watch on YouTube ↗
Takeaway

Shipping production AI agents requires the same eval-and-observability discipline Trainline applies via Braintrust to keep agentic ticketing reliable.

Summary

  • Hands-on workshop in London with Braintrust and Trainline (UK rail ticketing app) on shipping production-quality AI applications.
  • Trainline's AI engineers describe their move from BERT-era LLMs to state-of-the-art agentic products for ticketing and travel.
  • Workshop walks attendees through Braintrust's eval and observability tooling for managing agent quality at scale, with cheat sheets and Slack support.
  • Emphasizes that quality bars and continuous evaluation — not just shipping — are what separate prototype agents from production-grade ones.
evalsbraintrustagents
Original description
Getting a prototype working is straightforward. Making it reliable in production, especially with multi-step agents, tool use, and real users is the hard part. In this hands-on workshop, you'll work through the core parts of building production-grade AI applications with Giran Moodley, Mayank Soni, and Oussama Hafferssas.

Socials: 
- https://uk.linkedin.com/in/mayank-soni
- https://x.com/OussamaHaff
- https://www.linkedin.com/in/giran/

Timestamps

0:00 - Introduction and Welcome
4:07 - Workshop Overview and Agenda
4:39 - Understanding AI Engineering and Operational Challenges
9:55 - Introduction to Braintrust
12:56 - Experience from Trainline
28:35 - Building the Support Triage Agent (Overview)
33:57 - Basic Implementation: Single Shot Prompting
40:32 - Adding Local Tools for Determinism
41:30 - Implementing Specialist Stages (Agentic Flow)
46:19 - Instrumenting and Tracing the Application
56:43 - Evaluating AI Systems and Golden Data Sets
1:05:07 - Deploying and Managing AI in Production
1:13:58 - Online Scoring and Monitoring Production Logs
1:19:13 - Identifying and Remediating Failure Modes
1:33:05 - Key Takeaways and Summary
1:36:58 - Further Resources and Documentation