Matthew Deberry

Led the development of a FedRAMP-compliant, LLM-driven Generative AI platform (GPT-3.5, FLAN-T5, Vertex AI) for automating federal healthcare policy analysis. Architected multimodal document ingestion, semantic search, and summarization pipelines; delivered faster reviews, improved retrieval accuracy, and reduced hallucinations. Built OCR-based content extraction with LayoutLMv3; developed voice-enabled clinician interfaces; prototyped agentic reasoning for triage; implemented governance, monitoring, and reproducibility pipelines; ensured HIPAA/HHS compliance and secure, scalable cloud infrastructure.

Led the development of a major LLM-based Generative AI platform using GPT-3.5, FLAN-T5, and Vertex AI to automate federal healthcare policy analysis. Built multimodal document ingestion, semantic search, and summarization for 100K+ documents, enabling real-time responses during healthcare emergencies. Implemented a retrieval-augmented generation (RAG) pipeline with LangChain, FAISS, and Sentence Transformers to improve retrieval accuracy and reduce hallucinations. Developed a multimodal document understanding framework using Tesseract OCR, OpenCV, and LayoutLMv3 to extract and classify content from scanned policies, forms, and contracts, improving table and layout detection.

Led the development of a major high-impact, LLM-based Generative AI platform using GPT-3.5, FLAN-T5, and Vertex AI to automate federal healthcare policy analysis. Built multimodal document ingestion, semantic search, and summarization pipelines processing 100K+ documents, reducing manual review time by 56% and enabling real-time responses during the COVID-19 emergency. Designed a retrieval-augmented generation (RAG) framework with LangChain, FAISS, and Sentence Transformers to enable real-time semantic search, increasing retrieval accuracy by 30% and reducing hallucination-related errors by 22%. Implemented a multimodal document understanding framework using Tesseract OCR, OpenCV, and LayoutLMv3 to extract and classify content from scanned policies, forms, and contracts, improving table and annotation detection by 63%. Developed hands-free clinician interfaces via Whisper and WebSockets, enabling easier clinician workflow. Prototyped agentic reasoning systems for clinical escalation an