Chandrashekhar Medicherla

Title of the Talk: Index Selection as a Control Problem: Interpretable, Lightweight ML for IoT Vector Databases

Abstract:
Similarity search underpins a growing share of modern IoT applications, yet one of its most consequential choices, which index structure to use, is still left to human experts and rarely revisited as workloads change. This talk reframes that choice as an online control problem, and argues that the right tool is interpretable, resource-aware machine learning rather than a heavy black box. I’ll discuss why interpretability and small models often beat raw accuracy in the field, how to characterize a shifting workload well enough to act on it, and why the hardest failures tend to happen exactly where human operators struggle too. I’ll close by looking ahead, toward systems that learn online, share knowledge across organizations, and balance not just speed and accuracy but power and thermal limits at the edge. The goal is to leave the audience with open questions worth chasing, not just a finished result.

Bio

Chandrashekhar Medicherla is Lead Software Engineer – Database Infrastructure at Salesforce Inc., where he manages enterprise-scale database systems serving millions of users worldwide. With 18+ years of expertise in database infrastructure and cloud computing, he specializes in building reliable systems that achieve 99.99% uptime while processing billions of transactions daily. Chandrashekhar serves as Vice Chair of the Bluffdale ACM Chapter. He holds Fellow status with IETE, Senior Member status with IEEE, and Distinguished Fellow status with the Soft Computing Research Society. He has authored 16+ published research articles indexed in IEEE and Google Scholar and contributed 60+ peer reviews for IEEE and Elsevier publications. His work spans database systems, cloud architecture, and AI infrastructure across SaaS, financial services, healthcare, and internet-scale environments.