Research question
What happens when a generative component produces output that matches the expected schema but changes the underlying meaning? The pipeline may accept, enrich, aggregate, and publish the result before anyone recognizes the error.
Core idea
Silent corruption is a propagation problem, not only a model-quality problem. The relevant unit of analysis is the path from generation to downstream decision. Controls must therefore observe semantic changes, lineage, confidence, and amplification across stages.
Public work
Silent Data Corruption in Distributed GenAI-Augmented Data Pipelines: An Empirical Study of Hallucination-Induced Failure Propagation, with Sanjay Lokula. Presented in the IEEE ICDCS 2026 industry program.
Why it matters
Schema validation is necessary but no longer sufficient. As AI-generated content enters operational systems, semantic controls need to become part of the data path rather than an external review step.
Where the work goes next
Future work includes measuring how corruption changes under retries, model replacement, multi-agent handoffs, retrieval quality, and human correction—and identifying where a small intervention can stop the largest downstream impact.