Lan An Email & Phone Number

Los Angeles, California, Us

Global Ads Infra - Machine Learning Data Platform

Palo Alto, California, Us

• Implemented large scale ads recommendation training infrastructure from scratch and improved functionality, robustness, efficiency to support multiple models using TensorFlow.• Maintained, upgraded and developed end-to-end ads infrastructure for better use, including data pipeline, feature selection, model migration, mixed precision etc.• Applied MapReduce jobs to rank raw ads data, generating data process pipeline to be used in model training.• Built machine learning models based on ads recommendation system for ads ranking and pricing.• Implemented TensorFlow bottom layer operators including FFM, FwFFM algorithms for better performance.

China, China, Cn

• Built, trained, tested and refined fully convolutional network model using TensorFlow to calculate image white balance illumination and conducted bad case analysis for better performance.• Reorganized and rebuilt data input pipelines to make data-loading process more flexible and efficient.• Applied uniform sampling to generate illumination ground truth to merge brand-new datasets with image ground truth into model training and testing.• Converted auto white balance model training into multi GPUs solution and conducted model simplification.

北京, 北京, Cn

• Designed a system including denoising, clustering, aggregating and correlation analysis to summarize millions of alerts generated from the actual bank system; resulting in 95% total amount reduction and 99% key alerts maintenance with the help of apriori algorithm and Python Multiprocessing.• Detected and analyzed the cause and time of alert storm through frequency item analysis and change-point detection through Python Ruptures.• Implemented a system to directly generate alert storm analysis pdf report from the input data through Python docx; used by 10+ bank organizations and increased efficiency of operations.

北京, Beijing, Cn

Distributed Optimization Algorithms Design for Multi-Energy Systems• Completed modeling of multi-energy system and internally optimized single-user multi-energy system.• Designed both asynchronous and synchronous distributed iterative algorithms based on the original dual decomposition to solve the energy optimization problem in multi-user multi-energy model.• Introduced virtual random clock and redefined iteration numbers to prove convergence and optimality of algorithm.• Simulated correctness and applicability of algorithm on a complex system based on Matlab YALMIP.

Providence, Rhode Island, Us

LEMS-SLAM: A Fast and Robust Visual SLAM System Based on ORB and KLT• Extracted and tracked features from input images through ORB, SIFT and corner feature detector.• Selected keyframes according to feature maintenance rate and evaluated quality of keyframes’ feature correspondences generated from different methods on Blender Dataset.• Used ceres-solver to optimize the scale ratio between three consecutive keyframes by minimizing reprojection error.• Generated recovered tracks according to input images and tested results on KITTI Dataset with accuracy of 95%.

北京, Beijing, Cn

• Improved cutting-edge research on 3D face reconstruction tasks and finetuned models for a better performance.• Separated identity and nonidentity faces to build two separate CNN models, merged them before fully connected layer to better output based on TensorFlow.• Tested Normalised Mean Error (NME) and 3D Root Mean Square Error (RMSE) of different models on 3D dataset BU-3DFE, Meln3D dataset.

Master'S Degree, Electrical And Computer Engineering

Bachelor'S Degree, Electrical Engineering And Automation

High School Diploma