Experience

Amazon

Jan 2025 – Present San Jose, CA

Defined the technical strategy for function-calling in Amazon Q; led a team of 3 scientists to ship multi-step agentic reasoning and orchestration across enterprise knowledge sources.
Established Amazon Q's post-training methodology across instruction tuning and DPO preference optimization, adopted as the standard across the Amazon Q org.
Led a team of 9 scientists on safety and alignment for enterprise LLM deployment; designed the guardrail framework (automated red-teaming, content guardrails, PII-aware generation, safety classifiers) used across Amazon Q and Bedrock.
Defined Amazon Q's LLM evaluation methodology spanning multiple benchmark dimensions (MMLU, IFEval, MT-Bench, HumanEval, safety, function-calling accuracy) with automated regression testing and LLM-as-judge pipelines.

May 2022 – Dec 2024 Redmond, WA

Designed and shipped multimodal foundation model architectures for AR/VR eye tracking, with SWIN-based Vision Transformer pipelines, knowledge distillation, and quantization-aware training for on-device deployment.
Built cross-modal knowledge transfer systems using CLIP and cross-attention mechanisms for VQA and image captioning, enabling zero-shot generalization across unseen domains.
Developed large-scale deep sequence recommendation models for production ranking and retrieval, incorporating Transformer-XL architectures with meta-learning for cold-start optimization.
Developed privacy-preserving training pipelines and federated learning frameworks for distributed model training across privacy boundaries.
Drove research on self-supervised pre-training, contrastive learning, and domain-adaptive strategies adopted across multiple product surfaces.

Huma AI

Jun 2020 – May 2022

Built end-to-end cancer detection pipelines combining hierarchical CNNs, attention mechanisms, and self-supervised pre-training on pathology images, deployed on AWS SageMaker for clinical-scale inference.
Developed differentially private synthetic data generation frameworks using convolutional GANs and diffusion models, enabling HIPAA-compliant model training (published in Information Sciences, 140+ citations).
Designed adversarial debiasing and causal inference frameworks (Double ML) for disease prediction, reducing diagnostic bias while maintaining clinical accuracy.

Virginia Tech / West Virginia University

Jan 2016 – Aug 2020 Blacksburg, VA

Published 15+ peer-reviewed papers in top venues (IEEE Access, Information Sciences, arXiv) spanning NLP, computer vision, generative models, and audio-visual learning — accumulating 1,000+ citations.
Developed 3D CNN architectures for audio-visual speaker recognition and lip reading, establishing new benchmarks (150+ citations, IEEE Access 2017).
Authored a widely cited NLP deep learning survey (380+ citations, arXiv 2020).
Created high-impact open-source projects (TensorFlow-World, lip-reading-deeplearning, speechpy) with 8,000+ combined GitHub stars, adopted by researchers and practitioners worldwide.

Stealth Company

Sep 2012 – Dec 2014

Designed and optimized real-time digital signal processing algorithms for embedded systems, including adaptive filtering and spectral analysis modules.
Built low-latency audio processing pipelines for production hardware with strict throughput and memory constraints.

Work Experience