At Privacera, I engineered an AI-powered detection pipeline to identify PII, credentials, and confidential data across customer APIs, object storage, and LLMs. Semantic embeddings, contextual risk scoring, and policy-driven remediation helped reduce incident response time by 30%.
At the University of Maryland, I redesigned the Binoculars V2 LLM detector with out-of-distribution detection to address failures across unseen models and domains. I also engineered DeepSVDD and HRN on LLM embeddings, achieving 98.3% AUC.
At Protegrity Inc., I re-architected keyword search into a semantic embedding engine for enterprise data. Integrating Sentence Transformers with FAISS improved log-search accuracy and reduced false matches by 33%.
At Sanda Heathcare, I built a RAG-based medical prediction system on HIPAA claims and deployed a claim-validation pipeline that reduced denials and processing time by 35%. My projects include automated radiology report generation using PubMedBERT and a voice-and-chat expense-tracking application.
