At Amazon, I shipped a Java allowlist filtering layer in AVTransactionsEligibilityService that made the service privacy-safe by default and addressed PII leakage risk across 11 downstream microservices.
I also diagnosed and documented vulnerabilities in logging structures, coordinated audits of 19 API responses, and engineered distributed per-caller rate limiting with runtime-tunable TPS across an AWS ECS fleet.
For EffNetSwin, I designed a dual-stream classifier combining EfficientNetV2-S and Swin Transformer. It achieved 98.5% binary accuracy and 85% 5-class accuracy, with Grad-CAM overlays for interpretability.
I also built a RAG customer support assistant that searches PDF documentation and provides grounded answers with inline citations. Its pipeline uses embeddings and semantic search over ChromaDB, with a FastAPI backend and React.js interface.
