A production-oriented LLM/SLM system for local-service search: structured extraction, POI grounding, LLM-as-a-judge verification, agent planning, RL-style post-training, and nearline cache-based deployment. It improved multilingual query understanding, increased cache reuse, and reduced online serving pressure under real search-engine constraints.
A multimodal video Transformer system for harmful-content understanding, combining video/audio/text contrastive pretraining, temporal sequence modeling, supervised fine-tuning, attention-based harmful-content localization, and PyTorch/TorchScript productionization. It contributed to substantial reductions in harmful-content prevalence, including a major reduction in political misinformation, while improving reviewer efficiency through temporal localization.
A generative and interactive medical AI project for annotation-efficient 3D segmentation, combining diffusion-based image-mask synthesis, scribble-supervised learning, attention/CRF regularization, and online human-in-the-loop adaptation. The system reduced dependence on dense expert labels while improving segmentation robustness under limited-label and real-world deployment constraints.