Biography
Lokesh Karanam is a Staff Software Engineer specializing in Machine Learning Systems for Regulated Cloud Data Platforms and Generative AI Operations, backed by over six years of industry experience. He currently works at RR Donnelley (Austin, TX) and previously held a role as a Software Development Engineer at Amazon Web Services.
His educational background includes an M.S. in Computer Science from the University of Missouri.
Lokesh’s career highlights feature significant original technical contributions, such as developing a large-scale, privacy-compliance deletion platform and implementing solutions that drove substantial cloud cost savings and platform efficiencies. He also led an AI marketing platform project with a pending patent. His scholarly contributions include multiple peer-reviewed publications in prestigious conferences like IEEE DeSE and AMIA, covering advanced topics from Intrusion Detection using deep learning to Continuous Anticipation of Acute Kidney Injury (AKI) in the ICU. Lokesh’s contributions have earned him recognition, including service as a Session Chair at the IEOM conference.
Abstract
Title: Adding Consent Lineage to Your ML Pipeline Without Rearchitecting It
Abstract: The EU AI Act now requires organizations to document training data provenance, but existing MLOps tools (MLflow, model registries, data catalogs) track datasets and experiments, not individual-level consent. This talk introduces ConsentML, an open-source Python library that adds consent-aware data lineage to existing training pipelines with a single decorator. I’ll walk through how it captures subject-level lineage in a hash-chained audit log, connects to Snowflake and PostgreSQL data sources with credential-free provenance, handles consent revocations by tracing affected models, and verifies the audit trail’s integrity. I’ll also share three classes of false-verification bugs we found during development where the verifier reported a clean audit on tampered data and how we fixed each one.
