Curriculum Vitae

University of Colorado Boulder

• Thesis: The Role of Temporal Dynamics in Machine Perception
• Advisor: Michael C. Mozer. Committee: Samy Bengio, Clayton Lewis, Orit Peleg, Bradley Hayes

University of Colorado Boulder

University of Arizona

Meta

• Building multimodal LLMs for large-scale recommendation systems: foundation model pretraining, RL-based post-training (GRPO/DAPO), chain-of-thought distillation, and reasoning trace generation for next-ad prediction.
• Designing novel architectures that decouple reasoning from retrieval in LLMs for improved ranking performance and interpretability.
• Previously led contextual AI for smart glasses in Reality Labs: multimodal goal inference from speech, vision, and interaction signals for ultra-low-friction user interfaces.
• AI modeling team lead from data collection and annotation through training, evaluation, and real-time deployment on device. Delivered MVP demos achieving 95%+ accuracy at sub-500ms latency.
• Pioneered AI-native research workflows: built agentic experimental pipelines that autonomously orchestrate compute infrastructure, launch distributed training runs, and synthesize results — compressing multi-day experiment cycles into hours.

Meta

• Developed multi-scale contrastive predictive coding models for forecasting human actions at multiple levels of temporal abstraction.

Google Research Brain Team

• Researched novel neural architectures and loss functions inspired by massive parallelism and time-continuous information propagation in biological neural systems.

Google Research Brain Team

• Developed cascaded ResNets with biologically inspired parallel dynamics for anytime predictions, enabling adaptive speed–accuracy trade-offs through stateful, time-evolving computation. Published at NeurIPS 2021.
• Introduced a temporal-difference training loss that improved efficiency and robustness over state-of-the-art anytime-prediction methods.

Microsoft Research

• Advanced speech enhancement and activity recognition via multi-modal fusion research, resulting in two publications (CVPR, ICASSP) and two US patent applications.

Microsoft Research

• Proposed cross-modal squeeze-excitation fusion mechanism (AV(SE)²) integrating visual cues with audio features for speech enhancement. Published at ICASSP 2020.
• Demonstrated that time-based gating across feature layers improves intelligibility while reducing model parameters vs. standard concatenation-based fusion.

Microsoft Research

• Developed real-time 3D object retrieval and pose estimation from 2D RGB images using multi-view deep metric learning and fully convolutional networks.

Programming

Machine Learning

Tools & Infrastructure

Peer-Reviewed Conference Papers

Journal Papers

Preprints