For years, infrastructure engineering has relied on one core assumption: systems behave deterministically. Given the same inputs, software produces the same outputs, failures are traceable, and operational behavior can be reasoned about through logs, metrics, and alerts. LLM-powered systems challenge that foundation entirely. As language models move beyond chat interfaces into agents, copilots, deployment workflows, automation platforms, and operational tooling, infrastructure is becoming capable of interpretation, reasoning, and autonomous action. Modern systems are no longer just executing instructions — they are generating decisions. This introduces a strange new category of engineering problems. In this talk, we explore how the rise of LLM-native systems is changing the meaning of reliability, observability, and operational trust. Unlike traditional systems, AI-driven platforms can remain technically healthy while producing semantically incorrect behavior. A service may have perfect uptime while making flawed decisions, triggering unintended actions, or confidently generating incorrect outputs. We’ll discuss: Why deterministic engineering assumptions break in AI-native systems The difference between infrastructure failures and semantic failures Reliability challenges introduced by agents and tool-calling workflows Why hallucinations are operational problems, not just model problems The limits of traditional observability for LLM systems Emerging ideas around evaluation, guardrails, and AI reliability engineering What the future role of platform and reliability engineers may look like in increasingly autonomous environments This session is designed for engineers, builders, and operators working with modern AI systems who want to better understand the operational realities of deploying intelligent systems beyond demos and prototypes.
Kaustubha V is a Solution Architect at Microsoft in the Silicon Cloud organization, where she works on scalable cloud platforms, AI-driven systems, automation frameworks, and developer productivity solutions. Her work focuses on building reliable infrastructure and intelligent systems for large-scale engineering workloads across Azure and multi-cloud environments. At Microsoft, she designs solutions that use machine learning to optimize compute resources such as memory, runtime, and CPU requirements. She has contributed to production-grade ML pipelines, telemetry platforms, benchmarking frameworks, and cloud-native automation services that improve operational efficiency and simplify engineering workflows. Kaustubha has strong experience in Azure, AWS, Google Cloud, Kubernetes, distributed systems, and secure cloud deployments. She holds multiple cloud and AI certifications and has earned the Kubestronaut distinction from the Cloud Native Computing Foundation. Beyond her engineering role, Kaustubha is active in AI research and technology communities. Her research interests include generative AI, machine unlearning, responsible AI, and bias mitigation in machine learning systems. She has authored research papers accepted at international workshops and conferences, including NeurIPS,ICCV, GHC ICLR, GHCI, IEEE and leading data science venues. She is also passionate about mentoring and community impact. Through initiatives such as Women Techmakers and Microsoft’s Code Without Barriers, she supports students and early-career professionals in learning cloud computing, AI, and software engineering.
