Postdoctoral Research Associate · Analytics & AI Methods at Scale · Oak Ridge National Laboratory
Emily is a Postdoctoral Research Associate in the Analytics & AI Methods at Scale (AAIMS) group at Oak Ridge National Laboratory, working at the intersection of agentic AI, LLM safety, and exascale high-performance computing for science.
Her current research designs and hardens agentic AI systems for DOE national-laboratory scientific workflows — from architecting safety-mediation frameworks like VISTAGuard, to building multi-agent LLM pipelines for real-time experimental analysis at the Center for Nanophase Materials Sciences, to scaling evolutionary architecture search for hybrid Transformer–Mamba–MoE language models across 16,384 GPUs on the Frontier exascale supercomputer.
She holds a Ph.D. in Data Science & Engineering from the University of Tennessee, where her thesis advanced differentiable neural architecture search. She serves as Secretary on the 2026 Executive Board of the Oak Ridge Postdoctoral Association.
Unified security and safety-alignment mediation for tool-using LLM agents operating across HPC, instruments, and federated scientific infrastructure.
Trustworthiness evaluation, hypothesis-generation reasoning models, and domain-specialized LLMs for materials science and astrophysics.
Large-scale evolutionary and differentiable NAS for efficient hybrid Transformer–Mamba–MoE architectures at exascale.
Pipeline and hybrid parallelism for multi-billion-parameter models on AMD ROCm (Frontier MI250X) and NVIDIA CUDA platforms.
A unified security and safety-alignment mediation framework for agentic AI in DOE scientific workflows. Capability-aware safety gates span prompt, tool, RAG/memory, code, HPC, instrument, and federation boundaries, with interpreter-level trust-tag propagation and a quarantined-LLM architecture for structurally isolated handling of untrusted content. Paired with SciAgentBench, a companion red-teaming evaluation pipeline.
An agentic AI framework for molten-salt thermophysical-property analysis built with a PydanticAI agent loop (planning, memory, tool-use), a FastMCP server fleet, retrieval-augmented reasoning over a curated ChromaDB corpus, and a Next.js frontend with human-in-the-loop credential elicitation for SLURM job submission.
A multi-agent LLM pipeline for real-time spectroscopic data analysis at ORNL's Center for Nanophase Materials Sciences. Eight LLM-backed agents distributed across Engineering, Analysis, and Knowledge MCP servers, backed by a 100K+ publication RAG corpus, YAML material knowledge bases, and a live-experiment integration dashboard.
A distributed evolutionary architecture-search framework for efficient hybrid Transformer–Mamba–Mixture-of-Experts language models, scaling to 16,384 GPUs on the Frontier exascale supercomputer.
A trustworthiness-evaluation framework for scientific foundation models, with reusable benchmarks across truthfulness, adversarial robustness, safety, and ethics dimensions.
Supervised fine-tuning and Group Relative Policy Optimization (GRPO) applied to reasoning LLMs for materials-science hypothesis generation. Includes a chain-of-thought dataset of 100K+ documents and bespoke reward functions and evaluation pipelines.
Thesis: Generalized Differentiable Neural Architecture Search with Scaling and Stability Improvements · Advisor: Dr. Steven R. Young · GPA: 3.95 / 4.00
GPA: 3.94 / 4.00 · Outstanding Student in Computational Science (2016–2018)