Mariano Semelman
Mariano Semelman is a Lead Data Scientist and Machine Learning Engineer with over a decade of experience driving AI innovation in the e-commerce sector. Currently at OLX, he develops AI-powered media solutions that leverage Generative AI and agentic workflows to streamline the platform experience for sellers. He specializes in bridging the gap between research and real-world application, building scalable, end-to-end AI systems from the ground up.
He holds a Master of Science in Computer Science from the University of Buenos Aires (UBA), where he contributed to research at the university's Applied Artificial Intelligence Laboratory. Beyond his core professional and academic work, Mariano is an active speaker and technical contributor. He is the organiser of a local chapter of the Barcelona MLOps.community (part of the Agentic AI Foundation)
Session
PydanticAI makes it trivial to build a basic agent: define a model, add tools, call agent.run(). But what happens when that agent needs to remember a 20-turn conversation, hide tools that are not relevant, wait for human approval before a destructive action, or survive a process crash mid-execution?
This talk builds a single customer support agent live on stage, starting from a 10-line hello world and progressively layering ten PydanticAI capabilities until it handles real production concerns: typed dependency injection, message history, dynamic tool descriptions, lifecycle hooks, deferred tool calls, the capabilities system, tool search at scale, conversation compaction, and durable execution.
Each topic has a runnable demo. By the end you will know not just how to build an agent, but how to make one that is observable, safe, scalable, and resilient enough to deploy.