Software Engineer
Meituan
–
AI Research & Engineering
PhD candidate · KAIST · Former Software Engineer at Meituan
Doctoral research in trajectory modeling and representation learning, with current projects in generative traffic simulation and adaptive sensing.
Meituan
–
Advanced Institute of Information Technology, Peking University
–
Korea Advanced Institute of Science and Technology (KAIST)
2023 – Present
Cho Chun Shik Graduate School of Mobility; advisor: Prof. Inhi Kim
Zhejiang University (ZJU)
2018 – 2021
Ocean Engineering
Nanchang University (NCU)
2014 – 2018
Physics
Developed ProChunkFormer for sparse GPS reconstruction. On Porto taxi data at 240-second sampling intervals, reduced runtime by 20.7% and road-network MAE by 25.1% versus an autoregressive Transformer.
Designed SDA to measure the contributions of historical motion, current state, and goal information. Evaluated LSTM and Transformer forecasting on Porto taxi and ETH pedestrian data.
Combined LLM-derived semantics with heterogeneous graph learning. Achieved 0.6766 F1 in Porto-to-Lisbon transfer without retraining, compared with 0.6968 F1 for a Lisbon-trained model.
Combined LLM pseudo-labels with structural embeddings for cross-taxonomy alignment. With 1% human-labeled data, achieved 51.3% accuracy versus 40.0% for direct supervised training.
Project scope: Infrastructure-guided AV coordination, mixed-traffic modeling, and digital-twin scenario validation.
Lab project details : Physical AI for Mixed-Autonomy Traffic
Project scope: Dynamic-object behavior models and data-linked virtual environments for automated-driving validation.
Lab project details : Autonomous-Driving Virtual Test Environment
Project scope: Deep-learning traffic analysis and prediction, mobility-data fusion, and network-level simulation.
Project scope: Shared-mobility demand forecasting and operational planning across shared transport and mass transit.
Lab project details : Shared Mobility and Mass Transit Operations
Project scope: Using large language models for evolving agent-based mobility planners in autonomous-driving applications.
Developing route-conditioned traffic generation on SMART/CAT-K with a frozen pretrained motion model and trainable path-encoding and FiLM modules. Implemented recorded-scene data adapters and closed-loop evaluation; policy-optimization experiments are ongoing.
Built a local GPS replay web workspace with heuristic adaptive reporting, Kalman receiver-state estimation, historical reconstruction, playback controls, decision logs, and JSON run export.