Cedric Renggli

I develop data systems for efficiently organizing, maintaining, and querying the data used by AI applications. My previous work includes scalable search over pretrained models in SHiFT, efficient in-database learning without full data shuffling in CorgiPile, and incremental maintenance of vector-search indexes in Ada-IVF. Building on these foundations, I am interested in systems that automatically transform evolving data, model outputs, and interaction traces into maintainable and queryable knowledge, with statistical guarantees on answer quality and uncertainty.

Emerging AI workloads are increasingly stateful and interactive, comprising sequences of model calls, retrieval operations, tool executions, and user interventions whose resource demands evolve over time. I design serving architectures, state-management mechanisms, and adaptive scheduling policies for these workloads. This agenda extends my earlier work on communication-efficient distributed learning in SparCML and robust execution across unreliable infrastructure in Distributed Learning over Unreliable Networks. My current interests include managing intermediate execution state (e.g., KV caches), predicting future resource requirements from interaction traces, and coordinating model inference and secure tool execution across heterogeneous hardware.

The observed performance of an AI system reflects not only model capability, but also data quality, benchmark construction, infrastructure, execution environments, and stochastic variation. I develop methods for identifying and quantifying these effects. My work has exposed systematic ambiguities and metric biases in Text-to-SQL evaluation, introduced automatic feasibility assessment through data-quality analysis in Snoopy, and established statistically rigorous testing methods for ML pipelines in ease.ml/ci. By disentangling model behavior from data, system, and measurement artifacts, this research enables more reliable comparisons and better-informed decisions about complete AI systems.