At Meta, I built silent failure detection and tracing APIs that show which models and signals drive each ad safety decision, replacing hours of manual debugging. I also redesigned Brand Safety decision storage for a distributed service ingesting 100+ TB/day, reclaiming 2,000+ TB.
I rearchitected the Human Review decision pipeline into one end-to-end workflow and built a framework for experimenting with and rolling out content sensitivity model thresholds. The framework reduced setup time 67% and improved rollout time from 13 days to about 10.
At Moderna, I built a sensitive data classification system across 400 TB of AWS storage, raising classified data from 10% to 80%. Earlier at Amazon, I built a merchant enrollment backend that shortened onboarding from two weeks to three days, and at Ushur I developed a ClinicalBERT model that raised precision from 45% to 85%.

