Reimagining AI Systems through the Lens of Data Systems

Seminar: 4:00 PM – 5:00
Interactive session with students & researchers: 5:00 PM – 5:30 PM

Register for both using the same link: https://forms.gle/9eZ7i6pzdVMAHvNz5

When applying AI to real-world data, the bottleneck is rarely training one model or building one agentic workflow. It is about rapidly experimenting and iterating at scale, spanning data representations, hyperparameters, and nowadays, retrieval and agent configurations. I approach this challenge through a data systems lens, where we reimagine the notions of higher-level abstractions, query optimisation, and approximate query processing to help democratize AI systems. This lens has led to multiple productive collaborations with domain scientists and companies for workloads spanning DL, polystores, and LLMs. This talk will give a tour of those projects and conclude with some open challenges I see in data management for AI in this era of agents.

First, Project Cerebro reimagined DL model selection as multi-query optimization to raise experimentation throughput while reducing runtimes and GPU costs. Via Project DeepPostures, Cerebro powered multiple public health studies on sedentary and sleep behaviors using large-scale accelerometer time series data. Second, Project AWESOME is building a declarative polystore system for unified analytics over graphs, relations, and text. By enabling novel cross-model query optimizations, it helps applications in social media analytics for political science and cybersecurity.Finally, at my startup RapidFire AI, we are reimagining such DB ideas for LLM applications by building a high-throughput experimentation engine for configuring RAG, agentic, and fine-tuning pipelines. It enables AI developers to reach better eval metrics more quickly while reducing token spend and/or GPU costs.