I am a machine learning engineer at Meta working on GenAI model and serving infrastructure for creative generation systems and agentic workflows. My work spans diffusion transformers, LLM serving, post-training and RL, multimodal video editing, robotics perception, and real-time AI tutors.
I have 6+ years across ML research and engineering, including 3+ years in GenAI, LLMs, agents, image/video generation, post-training, inference, and infrastructure. Previously, I worked on Text-to-Edit at TikTok, barcode-less package identification at Amazon Robotics, and real-time AI avatar tutors at Thinkverse as the 1st Founding Engineer.
I was a PhD candidate in Machine Learning at Northeastern University, advised by Raymond Fu at SMILE Lab / FuLab, and completed an MS in Computer Engineering there. I was also an ML visiting scholar at UC San Diego.
Applied research and engineering across production GenAI, multimodal video systems, robotics perception, and real-time AI tutoring.
2025 - Present
Building large-scale GenAI model and serving infrastructure for creative generation systems. Work spans CUDA/Triton sparse attention kernels for Diffusion Transformer inference, adaptive sparse attention for video diffusion models, vLLM-based benchmarking for dense and MoE LLM serving, disaggregated prefill analysis, and RL/post-training infrastructure with VERL, GRPO, multi-objective RL, tool calling, and LLM-as-judge evaluation for agentic workflows.
2024 - 2025
Post-trained a multimodal LLM for controllable end-to-end video ad creation. The system turns product context, free-form instructions, video understanding signals, and slow-fast temporal evidence into structured edit drafts for advertising workflows.
2023 - 2024
Developed deployment-oriented perception models for package identification in robotics environments, including real-time edge localization and optimized ONNX/TensorRT inference for barcode-less and barcode-based package understanding.
Amazon Science2023
1st Founding Engineer for an AI tutoring platform. Designed and built the end-to-end full-stack architecture for real-time avatar math tutors, moving student speech through ASR, LLM reasoning, voice generation, and lip-sync avatar rendering with user-provided character images as the base.
ThinkverseMy research started from domain adaptation and generalization, then moved toward multimodal generation, controllable video editing, and deployed AI systems.
I was a PhD candidate in Machine Learning at Northeastern University, advised by Raymond Fu in SMILE Lab / FuLab, and completed an MS in Computer Engineering there. Earlier research includes ML visiting scholar work at UC San Diego.
Domain adaptation, domain generalization, medical imaging robustness, semantic segmentation, multimodal video understanding, and LLM-driven controllable editing. The common thread is making learned systems work outside narrow benchmark assumptions.
Selected research outputs across multimodal generation, video editing, domain adaptation, and generalization.
2025
2024
Pseudo Label Self-Refinement · WACV
Paper
2024
US Patent Application 20240312197
Patent
2024
Four-Level Optimization · TACL
Paper
2023
DANN-DG · Machine Learning
Paper
2023
Source Data Reweighting · Scientific Reports
Paper
2022
Fourier Style Calibration · IJCAI
Paper
2022
2021
White Matter Hyperintensity Segmentation · ISBI
Paper
2021
PAC-Bayesian Medical Imaging
Paper
2021
Multi-Domain Meta-Learning · ISBI
Paper
2020
I am interested in building reliable multimodal AI systems, agentic workflows, and deployment infrastructure for frontier AI products.