Los Angeles Metropolitan Area
Algorithm and large-scale data processing expert with over 15 years of experience, specializing in Natural Language Processing, generative models, and computer vision. Experienced in the full lifecycle from research prototyping to production deployment. Skilled at optimizing complex algorithms and solving large-scale data challenges. Proven leadership experience managing engineering teams and translating research outcomes into production systems that significantly improve product performance and user experience, while fostering technical growth within teams.
Led a 5-person team developing the search engine question-answering system QAS-V2 (internal version v2.3), addressing data imbalance and long-tail query challenges. Improved system accuracy by approximately 14–16% while continuously optimizing query response time. Contributed to the core development of the RecomX personalized recommendation engine, improving click-through rate by about 10% through A/B testing and iterative feature engineering. Addressed cold-start user data scarcity that initially caused model bias. After three rounds of simulated user behavior experiments and loss function adjustments, model performance gradually improved. During mobile deployment, inference latency initially failed to meet targets. Introduced lightweight quantization strategies, reducing latency by 18–22% and supporting tens of millions of daily requests.
Conducted research on generative text and image models, proposing an improved architecture LayerNorm-Transformer (experimental ID: LN-TX19) that significantly improved performance in low-resource text generation tasks. Used distributed training and mixed-precision optimization, reducing training time by around 30% and increasing experimental efficiency. Early models suffered from slow convergence on complex generation tasks. Stability was achieved after experimenting with multiple optimizer combinations and learning rate scheduling strategies. Published two papers in NeurIPS/ICML, and maintained well-documented experimental logs and reproducible code repositories.
Developed the GPU-accelerated image recognition model VisionX (v1.1) for industrial inspection and autonomous driving prototypes. Improved model accuracy by approximately 18–22% using transfer learning and data augmentation on small datasets. Early models showed high false-positive rates in complex scenarios, temporarily causing performance regression. This was gradually improved through multi-scale input strategies and regularization techniques. Responsible for model training pipelines, deployment scripts, and experiment logging for a 3-person engineering team.