Hengyi Tang

Experience

• 

  Jiangsu Advanced Memory Semiconductor Co. Ltd.

  Aug 2021 - Feb 2022

  Machine Learning Engineer

  - Developed a rapid preliminary screening system for wafer lithography yield using intelligent data processing via Ali Cloud, enabling efficient reading of raw data and writing of supplementary information.- Designed and integrated a hybrid model combining Resnet-50/101 and LSGAN to analyze electron microscopy photos, accurately detecting wafer lithography depth and regularity, and marking typical problem areas to improve manual screening efficiency.- Implemented performance optimization strategies using tools like Java Profiler and SQL Optimizer, significantly enhancing the system’s ability to manage datasets exceeding 1 million records. Show less

• 

  Veho

  May 2023 - Aug 2023

  Machine Learning Engineer

  - Developed an LSTM model in a Docker-based environment to predict urban road short-term traffic flow in San Francisco and infer future road capacity and potential damage risks, utilizing SQL to extract 3,000,000 data from PostgreSQL over 5 years. This led to a 12% improvement in logistics efficiency, optimizing route planning and reducing delivery times.- Constructed a ViT-SRGAN hybrid model with Pytorch and OpenCV for surveillance image super-resolution and fine-grained license plate recognition, with real-time inference in the truck information database deployed via TensorRT.- Advanced pre-processing and feature engineering: Leveraged XGBoost for feature augmentation by fine-tuning hyper-parameters. Integrated a custom Butterworth filter using SciPy to reduce noise by 60%, improving signal clarity in time-series data. Enhanced the K-Means clustering algorithm in Scikit-learn for superior outlier detection. Show less

• 

  USC Information Sciences Institute

  Aug 2023 - Nov 2024

  Research Assistant

  - Developed an adaptive, robust edge system for multi-models on distributed edge devices, ensuring model accuracy while reducing communication and inference costs by 8%-23%.- Implemented PipeEdge, a distributed framework using pipeline parallelism strategy in Pytorch and Colossal-AI that fine-tunes traditional CNN (Resnet, VGG, Alexnet) and transformer-based models (ViT, BERT, DEIT, LLaMA) to enable arbitrary layer sharding for efficient distributed processing across edge devices.- Employed AdaptivFloat, a dynamic float-32 precision quantization technique that autonomously clips the dynamic range of layer outputs and applies exponential-scale non-uniform quantization, minimizing the error of frequently occurring values and providing faithful quantization mapping from int-4 to float-8 under limited channel capacity between edge devices. Show less

• 

  UC Berkeley Cyber-Physical System Design Lab

  Jun 2024 - now

  Research Assistant

  - Developed a multi-model system in PyTorch Geometric and DeepRobust with weak-to-strong generalization that utilize the text understanding capabilities of LLMs for feature augmentation in Text-Attributed Graphs (TAGs), effectively improving the robustness of GNN models (GCN, RevGAT, GAMLP) against structural perturbations by over 18%.- Supervised GNN training on clean and attacked graphs by incorporating last-layer hidden states and predictions of LLMs into feature vectors, generating pseudo labels for test data from LLMs, and modifying the loss function. Show less

• 

  USC Viterbi School of Engineering

  Sept 2024 - Dec 2024

  Computer Science Grader
• 

  Skyline Dev Labs

  Mar 2025 - now

  Deep Learning / AI Engineer

  LLM model fine-tune, dataset preparation, and prototype AI agents using prompts and ML algorithms.Work for Savetolater AI project.