Yulia Stolin

AI (ML) Platform Manager
Teads
Israel

About

Yulia Stolin leads the ML Platform at Teads, building production-grade infrastructure for data science research, experimentation, and large-scale inference.She has more than 20 years of experience across software architecture, backend systems, big data, machine learning, and data platforms. Her work sits at the crossroads of backend engineering, DevOps, data engineering, data science, and ML, where architecture must meet real-world constraints around scale, latency, reliability, cost, and business impact.Outside work, Yulia enjoys traveling, playing tennis, and observing how culture, technology, and people interact, miscommunicate, adapt, and occasionally create something surprisingly useful together.
Talk

Yulia Stolin | From AutoML to AgenticML: A True Paradigm Shift in ML Platforms

AI Engineering, Agentic AI, ML Platform, Data Science
AI agents are not only changing how teams write code. They are changing how machine learning research is structured, executed, and scaled.In this session, Yulia Stolin will share a production-oriented journey from automated machine learning (AutoML) to AgenticML: a shift from predefined automation toward AI-driven research loops where data scientists, ML platforms, and AI agents work together.Yulia leads the ML Platform team at Teads. The platform serves the research needs of the company’s data science groups and supports large-scale production ML systems, including inference infrastructure processing more than one billion predictions per second. In this environment, every promising idea must eventually satisfy requirements related to scale, latency, reliability, infrastructure cost, reproducibility, and business impact.Historically, AutoML helped machine learning teams automate known research paths such as feature selection, hyperparameter tuning, architecture search, model training, and repeatable experimentation. AutoML made research faster, more structured, and more scalable.But AutoML works mainly inside boundaries that already exist. It can search, tune, and optimize inside a space that humans have already defined.The next step was adopting the Model Context Protocol (MCP): giving agents access to context, connecting them to platform APIs, and allowing them to automate parts of the research workflow. This made agents much more useful because they could interact more directly with the real system by reading metadata, triggering actions, accessing tools, and executing experiments.But this was still a closed system. Agents could automate only what the platform already exposed, but the research was still limited by predefined tools, predefined flows, and predefined assumptions.AgenticML represents the real paradigm shift: from a closed, predefined automation system to an open research system that can reason about the next step, decide what is missing, implement what is required, run experiments, inspect metrics, compare results, and continue working across long-running research sessions.The session will focus on practical implementation. It will show how the platform evolved from predefined automation to an AI-first research approach and what this shift requires in terms of tools, workflows, platform capabilities, and engineering mindset.The talk will include real-world use cases, implementation examples, and successful projects, while also covering the challenges that arise when agentic AI moves from demos into production-grade research platforms: experiment execution, long-running agent sessions, cost control, GPU usage, reproducibility, platform safety, decision tracking, and agent-friendly reporting.Attendees will leave with a clear understanding of the AgenticML paradigm shift, why it matters for ML platforms, and how backend engineers, ML engineers, data scientists, and platform teams can approach building the next generation of AI-driven research systems.