I am excited to share that I have accepted a position at Amazon, where I will work on silent data corruption and reliability of systems for training/inference at scale.
Silent data corruptions are hardware errors that go undetected: a faulty chip returns a wrong result without raising any alarm, and the error propagates into the application’s output. At fleet scale, even rare defects show up regularly, which makes screening for them and measuring how well tests cover them a practical problem. My work will focus on methodologies for both.
The role is a natural continuation of my Ph.D. research at Columbia, which I am finishing this spring, and where reliability became a central part of my work.