Joonhun Lee

Qraft Technologies

Leading AI research across order execution, market and asset class expansion, LLM-mediated financial signaling, and ML-driven asset allocation — translating research-stage agentic patterns into deployable decision systems under real-world constraints.

Built and deployed reinforcement learning systems for order execution and futures trading, designing policies and evaluation pipelines robust to noisy, non-stationary, and partially observable market environments.

Wavebridge

Engineered trading infrastructure and quantitative research systems for digital-asset market making — supporting systematic strategy research and decision analysis under realistic execution constraints.

Doctor Now

Drove Office of the CEO workstreams across investor relations, business development, operations, and data/product initiatives during a period of rapid growth — bridging product, data, and operations into actionable decisions.

Seoul National University

CFA Institute

Seoul National University

Generalized Gaussian Temporal Difference Error for Uncertainty-aware Reinforcement Learning

Uncertainty-aware reinforcement learning has long assumed TD errors are zero-mean Gaussian, which miscalibrates how confident an agent should be in its own value estimates. A generalized-Gaussian replacement — with a learnable shape parameter governing uncertainty and a kurtosis-aware sample weighting — improves sample efficiency for both SAC and PPO across MuJoCo continuous control.

Feature-aligned N-BEATS with Sinkhorn divergence

Time-series forecasters tend to collapse when the test distribution drifts — an unseen market regime, a new region's weather — because they have no mechanism for learning what is invariant across source domains. Stack-wise optimal-transport (Sinkhorn) alignment of N-BEATS feature distributions across source domains restores generalization under severe shift, while preserving the model's interpretable basis decomposition.

MINR: Implicit Neural Representations with Masked Image Modelling

Masked autoencoders inherit the masking strategy they were trained on — change the mask shape or move to an out-of-distribution image and reconstruction collapses. MINR reframes masked image modeling as predicting a continuous coordinate-MLP from the visible pixels via a transformer hypernetwork, reconstructing across unseen masking patterns and OOD domains with roughly 7x fewer parameters than MAE-Large.

MCSA (SNU Management Consulting Student Association)