I make ambitious AI ideas work fast, reliably, and at scale.
Over the past decade, I have built production ML systems across finance, healthcare, and manufacturing while exploring reinforcement learning for imperfect-information games and competing on Kaggle.
Since 2012, I have tackled dozens of industrial and academic ML benchmarks on Kaggle and beyond. Several winning solutions led to CVPR workshop papers and open resources, including the GLDv2-clean dataset later used to train DINOv2.
Through these competitions, I became a 5× Kaggle Grandmaster, earning 25 gold medals and reaching a career-high global rank of 4th, and served as a Google Developer Expert for Kaggle (2025–2026).
Hidden information, long action sequences, and interacting card effects make Pokémon TCG a challenging RL environment. Our team trained a Transformer policy with behavior cloning and 273 million games of CUDA-accelerated PPO self-play, then fine-tuned matchup specialists with expert replay review.
I develop RiichiEnv, a fast Rust simulator with a Python API for reinforcement learning, and run RiichiLab, an open arena for competing mahjong bots. Together they support training, online matches, and replay analysis.