Please note: This PhD seminar will take place in DC 3301.
Xiangru Jian, PhD candidate
David R. Cheriton School of Computer Science
Supervisor: Professor Tamer Özsu
Foundation models make it possible to ask data systems questions in natural language, but the prompt is only the front door. A single analytical request may span relational databases, knowledge graphs, document collections, and visual data, each with its own native operations, cost profile, and failure modes. The right source may become clear only after an intermediate result arrives, the steps are dependent, so errors compound along the chain, and a wrong answer is useless unless it can be traced back to the step and source that produced it. A useful interface must therefore do more than understand intent. It must coordinate the native operations of heterogeneous systems and return a result that can be checked.
This talk presents an architecture built on three ideas. First, no single model should speak for every source: each system deserves an agent that understands its language, or, where none exists, defines one. Second, every agent should commit to an explicit, inspectable intermediate object, such as a query or execution plan, that the underlying system can run and validate before the next step proceeds. Third, an answer should stay connected to the evidence behind it, so that a failure can be located rather than merely detected. Together, these ideas position foundation models as a semantic interface to data systems, not a replacement for their query engines. I present results for the source and interface layers and close with the open problem I am currently working on: an orchestration layer that discovers what each source can answer, composes dependent operations across them, and adapts when intermediate results change the task.