resident researcher | remote

• exploring novel dual-stream reinforcement learning architecture that addresses fundamental limitations in current rl research through multi-scale temporal hierarchies and physics-aware learning, contributing to advances in safety-critical robotics simulation.
• investigating multi-algorithm comparison framework for systematic evaluation of conventional vs dual-stream reward schemes across dqn, ppo, sac, and a3c, with real-time genesis physics simulation and taichi backend integration for authentic robotic control.
• contributing to production-ready rl systems through modular environment adapters, automated training pipelines with progress tracking, and comprehensive model management for physics simulation and game environments, aiming to advance safe ai.

research associate | collaboration: university of birmingham & u.s. army research institute

• developed a heterogeneous-agent reinforcement learning framework using human-proxy agents that simulate realistic human constraints and capabilities to improve ai-human collaboration in multi-agent systems.
• designed a cooperative grid-world capture environment based on stag hunt game theory, where machine agents had full observability but couldn't detect target health, while human-proxy agents had limited vision but unique disease detection abilities.
• conducted experiments across various environment configurations, varying disease probability and penalty severity to analyze cooperation patterns.
• demonstrated that rl agents trained with human-proxy teammates achieved superior cross-environment performance, with teams trained under moderate risk conditions showing 30-40% higher collaboration rates.

ai ml intern | san jose, remote

• built a deep learning pipeline with tensorflow and pytorch using resnet-50 and transformer models for classification. used tensorrt to optimize inference, achieving 25% faster performance and 15% higher accuracy.
• created distributed ml infrastructure with apache airflow and mlflow on aws gpu clusters using horovod, reducing training time by 20%. set up automated data pipelines for feature engineering.
• developed a predictive analytics system using postgresql and bigquery with scikit-learn and xgboost on imbalanced data. optimized queries to improve decision-making efficiency by 30%.