Los Angeles, California, United States
PhD in Electrical and Computer Engineering from UCLA, focused on applied machine learning and generative AI. I work on transformer-based architectures, diffusion models, vision–language models, and multimodal learning, with experience building scalable ML systems that translate research ideas into real-world deployment. I’ve worked across video, text, and time-series data, designing end-to-end pipelines for representation learning and foundation model fine-tuning, with a focus on practical, reliable deployment. I combine rigorous research training with an applied mindset, and I’m motivated by building high-impact ML systems in fast-paced, collaborative environments.
• Designed and implemented an end-to-end multimodal AI pipeline for video-based event detection using vision-language models, demonstrating generalizable strategies for time-series and video data. • Applied advanced prompt engineering techniques to enhance temporal reasoning in foundation models (Gemini), contributing to more accurate and explainable outputs. • Developed a fine-tuning dataset via anonymized video preprocessing and human-in-the-loop annotation refinement. • Fine-tuned open-source multimodal models for task-specific video-text alignment, optimized for local deployment and interpretability.
• Built a generative AI framework for conditional 3D human motion synthesis using VAE and contrastive learning, applicable to animation, robotics, and behavior modeling. • Designed a hierarchical decoding strategy that improved trajectory reconstruction and downstream motion prediction. • Achieved performance gains over existing benchmarks (HumanAct12, UESTC, MoCap), validating the model's generalizability and robustness.
• Leveraged transformer models to decode human conceptual thinking patterns from intracranial EEG data across diverse brain regions. • Developed a novel multi-channel cross-attention transformer architecture to integrate multimodal neural signals. • Achieved concept decoding seconds before vocalization, revealing neural reinstatements during memory recall. • Conducted lesion studies linking decoding accuracy to specific brain regions, enhancing model interpretability.
• Designed an agentic pipeline to convert NHTSA crash reports into simulation-ready driving scenarios • Built OSM map-aligned trajectory synthesis enforcing strict kinematic and spatial constraints • Developed a simulator integration toolchain for converting structured trajectories into prototxt, 2D replays, and CARLA 3D scenes, with diffusion-based (NVIDIA Cosmos-Transfer) cinematic rendering • Implemented validation and retry loops for geometry, execution, and downstream compatibility • Designed hybrid evaluation using programmatic checks and LLM-as-judge rubric scoring for quality
• Delivered production-ready ML solutions for urban smart parking systems, covering the full pipeline from data acquisition to inference and model optimization. • Fine-tuned object detection and tracking models for real-world robustness using YOLOv3, optimized for diverse lighting and occlusion conditions. • Filed two patents for innovations in visual tracking and parking automation technologies.