I am an AI Scientist at Pacific Northwest National Laboratory (PNNL) in Seattle, WA. I received my Ph.D. in Computer Engineering from Duke University, supported in part by a DOE-SCGSR Thesis Fellowship.
My research centers on computationally efficient, geometry-aware generative modeling, guidance for diffusion and flow matching, and monitoring/evaluation of foundation models.
Demonstrates that linear probes on intermediate LLM activations reliably catch semantic tool-calling failures across 18 models on the Berkeley Function Calling Leaderboard and generalize to unseen error categories.
Develops an automated VLM/LLM-powered framework to audit concept unlearning in text-to-image models, achieving a ~10% improvement in targeted knowledge removal.
Introduces an enhanced reference-free pre-training data detection scoring metric for LLMs, delivering substantial gains in membership inference AUROC.
Introduces training-free unconditional guidance via shifted power iterations that avoids saddle regions, reducing EDM2 Fréchet distance on ImageNet by 40% at half the memory cost of CFG.
Proves denoising score matching loss is lower-bounded by data intrinsic dimension, enabling a scalable estimator which significantly cuts GPU memory and latency with 25% lower MAE.
Proposes Gaussian Channel Autoencoders (GCAE) to overcome parametric Gaussian limits in representation disentanglement via Dual Total Correlation.
Derives scalable game-theoretic adversarial losses for autoencoders to produce decorrelated, information-rich latent spaces without explicit factor supervision.
Leverages complex-valued network layers to resolve standard objective trade-offs during gradient regularization, raising PGD adversarial accuracy by 10%.
AI Scientist · Seattle, WA
Lead research on continuous-time generative models and trustworthy AI systems. Developed Saddle-Free Guidance, a geometry-aware, training-free unconditional sampling method matching CFG cost while cutting unconditional ImageNet Fréchet distance by 40% with EDM2. Formulated a theoretical framework showing Denoising Score Matching loss bounds local intrinsic dimension (LID), yielding a SOTA LID estimator that requires significantly less GPU memory and latency. Designed automated VLM-driven evaluation suites to audit concept unlearning in foundation models, boosting targeted unlearning fidelity by ~10%, and built representation-based LLM tool-call verification achieving 90% AUROC in tool-call error prediction.
DOE SCGSR Fellow · Oak Ridge, TN
Conducted thesis fellowship research focused on geometric deep learning and AI for scientific discovery. Designed and implemented 3D equivariant graph neural network (GNN) diffusion architectures in PyTorch Geometric for atomistic materials generation, demonstrating conditioned spatial coordinate and physical property generation on the QM9 benchmark.
Ph.D. Research Intern · Remote
Researched robust representation learning strategies for vision models in adverse environments. Developed novel representation objectives for object detectors that increased corrupted-input mAP by ~30% while maintaining performance on clean distributions.
Ph.D. Researcher & Teaching Assistant · Durham, NC
Authored doctoral dissertation on regularization methods for interpretable, robust deep representations and trustworthy systems. Co-authored Min-k%++ for reference-free LLM pre-training data detection (ICLR '25 Spotlight), introduced adversarial covariance minimization for latent disentanglement (ECCV '22, ICLR '23), and engineered complex-valued neural network layers improving adversarial robustness under gradient regularization (ICML '21). Served as graduate teaching assistant for undergraduate and graduate computer architecture.