Keynotes

Immanuel Trummer

Associate Professor of Computer Science

Cornell University

Towards Efficient Semantic Query Processing Engines

Abstract: Large language models (LLMs) have expanded the scope of databases from purely relational to semantic query processing engines. These engines support SQL queries with AI operators on multimodal data, configured via natural language instructions, that process functions requiring deep semantic understanding of data. Semantic query processing has been adopted quickly in industry. However, the high overheads of LLM invocations limit its scalability.

In my talk, I will focus on methods that make semantic query processing more efficient. First, I will present SemBench, the "TPC-H of semantic query processing", enabling us to benchmark cost-quality tradeoffs realized by semantic query processing engines. Second, I will discuss our work on ThalamusDB, leveraging ideas from approximate query processing and specialized semantic operator implementations to speed up semantic queries. I will conclude by describing ongoing projects and avenues for future research.

Bio: Immanuel Trummer is an associate professor of computer science at Cornell University. His research focuses on making data analysis more efficient and more user-friendly, often leveraging techniques from the area of machine learning, in particular, large language models. His papers were selected for "Best of VLDB", "Best of SIGMOD", and the CACM Research Highlight Award. He received an NSF CAREER grant for his work on database tuning via LLMs and multiple Google Faculty Research Awards. He is also the author of the book "Data Analysis with LLMs", now available in five languages.

Balakrishnan (Murali) Narayanaswamy

Principal Machine Learning Scientist

Amazon Web Services

AI Agents as Data-Processing Systems? Learnings from Production Data Systems

Abstract: AI agents are shifting from producing answers to changing systems. They can already diagnose incidents, retrieve enterprise knowledge, generate SQL, optimize resources, and invoke tools that alter shared state. Many of the failures that erode trust arise not in the model but in the surrounding harness: choosing the wrong source, hiding a legitimate ambiguity, repeating reasoning it could have reused, asking the model to validate its own error, acting under unclear authority, or retrying a partial action and compounding the damage. These are not failures of intelligence; they are failures to carry the user's implicit intent through to the action it becomes. Learning from how we use schema and other ideas to manage structural chaos over data sources - what if we view an agent not merely as a model with tools, but as a repeated use, stateful data-processing system with a stochastic planner inside?

In this talk, I will draw on lessons from deploying ML- and AI-based systems in production across AWS, including learned workload management, query-time prediction, and scaling in Amazon Redshift; security monitoring under drift and troubleshooting in Amazon RDS; and grounded LLM assistance, ambiguity-aware Text-to-SQL, and context graphs for enterprise agents. Across these systems, useful predictions became deployable only inside machinery that supplies explicit objectives, reliable performance, bounded decision domains, fallbacks, independent checks, permissions, and feedback. Viewed as a data system, production agent surfaces familiar tensions, each of which we may be able to address with familiar mechanisms: reuse versus generalization, exploration versus exploitation, abstention versus action, and provenance versus recovery, handled variously by declarative workflows, general-purpose agents, formal methods, human review, logs, transactions, and branches, which provide different guarantees and leave different gaps. Rather than attempt to crown one paradigm, I will ask how these mechanisms should compose, and where, if anywhere, the data system should end and the agent system begin.

Bio: Balakrishnan (Murali) Narayanaswamy is a Principal Machine Learning Scientist at Amazon Web Services, where he leads a team in AWS AI. His research lies at the intersection of AI, ML, systems, optimization, learning, and inference. He has led the science teams behind Amazon Redshift ML, DevOps Guru for RDS, and GuardDuty RDS Protection, and, more recently, AI-driven scaling and optimization in Redshift Serverless, Amazon Q Generative SQL, Structured Knowledge Bases in Amazon Bedrock and built agents Amazon Quick. His work on Amazon Personalize received the Best Paper Award in the Applied Data Science Track at KDD 2020, and their recent research on ML for systems appears at venues including Neurips, ICML, AAAI, UAI, SIGMOD 2023 and 2024 and VLDB and CIDR 2025.

Page updated

Google Sites

Report abuse