Agentic Systems at Scale with Vilva Athiban (Omio)

About this episode

What does an agentic AI system look like when it has to work 100% of the time for millions of travelers, not just in a demo? Vilva Athiban is Lead AI Engineer at Omio, one of Europe's leading travel platforms. In June 2024, Omio picked him and one other engineer to start an AI team reporting directly to the CEO and CTO. He had no ML background. All he knew was how to call an LLM API. Since then he has built Omio's internal AI platform, the MCP integrations that connect agents to company systems, and Omio.ai, a consumer-facing agent product built entirely on MCP with no agent framework. In this episode Vilva walks through the full evolution of the architecture: LangChain, then LangGraph for control, then no framework at all. He explains the difference between supervisor and orchestration agents, why bigger models turned out cheaper than mini models at scale, how they cap tool calls and cache prompts to control cost, why guardrails ended up with the legal team, and why he thinks loop engineering will die in a few months. We also cover harnesses and context compaction, whether RAG is dead, how to pick models in 2026, fine-tuning, and what a JavaScript engineer should learn to move into AI. Key Topics:- From JavaScript engineer to Lead AI Engineer with no ML background- LangChain vs LangGraph, and why Omio's backend is now 100% MCP- Supervisor agents vs orchestration agents- Bigger models with medium thinking were cheaper than mini models- Capping tool calls, summarizing tool responses, prompt caching- Guardrails are subjective: the whale hunting debate- The agent is only as powerful as the tools you give it- What MCP is and why it replaced RAG for Omio- Harnesses, memory, and context compaction in Claude Code- Loop engineering: a sugar-coated while loop- How to choose models in 2026 (big, mini, nano)- Why fine-tuning is coming back in 2027- Are developers cooked? The autocomplete that writes 15 files- How JavaScript engineers can become AI engineers- The token bubble, budgets, and the move to open source models CONNECT WITH VILVA💼 LinkedIn: https://www.linkedin.com/in/vilvaathiban/🐦 X: https://x.com/vilvaathibanpb🐙 GitHub: https://github.com/vilvaathibanpb🌐 Website: https://www.vilvaathiban.com/ FOLLOW & SUBSCRIBE🎙️ Podcast: https://neciudan.dev/senors-at-scale📧 Newsletter: https://neciudan.dev/subscribe💼 Dan on LinkedIn: https://www.linkedin.com/in/neciudan💼 Señors at Scale on LinkedIn: https://www.linkedin.com/company/senors-scale/📸 Señors at Scale on Instagram: https://www.instagram.com/senorsatscale/📸 Dan on Instagram: https://www.instagram.com/neciudev ADDITIONAL RESOURCES- Anthropic Learn: https://www.anthropic.com/learn- Model Context Protocol: https://modelcontextprotocol.io- Omio: https://www.omio.com #AIEngineering #MCP #AgenticAI #LLM #LangChain #LangGraph #SoftwareEngineering #AIAgents 💬 Have you hit the point where a bigger model was cheaper than a mini model in production? Tell me in the comments.