Software Engineer, Machine Learning at Google DeepMind
AI is analyzing your overall score…
Identifying your key strengths…
Evaluating your skill match against the job requirements…
Assessing your cultural and operational fit
• 10+ Years Industrial Experience on Machine Learning. • 20+ Years Research Experiences on Machine Learning. • Rich Research and Industry Development Experience on Software Systems.
Chinese Academy of Sciences
Doctor of Philosophy (Ph.D.), Computer Vision and Machine Learning
January 1, 2007 – January 1, 2010
Shandong University
Master's degree, Pattern Recognition and Digital Image Processing
January 1, 2004 – January 1, 2007
Google DeepMind
Software Engineer, Machine Learning
April 1, 2022 – Present
Greater Seattle Area · On-site
Amazon Web Services (AWS)
Machine Learning Engineer
August 1, 2021 – April 1, 2022
Greater Seattle Area · On-site
Career Break
Full-time parenting
June 1, 2019 – August 1, 2021
Greater Chicago Area
The Home Depot
Team Lead
March 1, 2015 – June 1, 2019
Greater Atlanta Area · On-site
Sears Holdings Corporation
Research Engineer / Development Manager
June 1, 2013 – February 1, 2015
Greater Chicago Area · On-site
Oregon Health & Science University
Senior Research Associate
June 1, 2010 – May 1, 2013
Portland, Oregon Area · On-site
Bard Personalization Evaluation
February 1, 2024 – Present
Led the design and end-to-end implementation of Bard’s personalization quality evaluation framework, now deployed in production to continuously monitor personalization performance. The system integrates Autorater pipelines with human evaluation workflows to assess key principles (Appropriateness, Groundedness, Content Quality, and Style) achieving over 80% alignment between Autorater and human ratings across all dimensions. Developed and scaled privacy-safe No-Look Evaluation (NLE) pipelines, supporting both Owned Test Accounts (OTAs) and consented real-user data for large-scale evaluations. Introduced and optimized Autorater prompt-tuning strategies, improving sensitivity to nuanced personalization behaviors and enabling scalable and repeatable quality tracking that directly informs Bard model updates.
User Activity Prediction
June 1, 2023 – February 1, 2024
Designed and implemented an LLM-based user activity prediction system to forecast user actions (such as visits, searches, or video watches) by modeling temporal and spatial behavior patterns. Fine-tuned Transformer-based models that improved Top-5 recalls 15% over baselines and introduced custom evaluation metrics (category and location match) to capture near-misses aligned with user intent. My work improved the contextual relevance and personalization accuracy of Google’s location-based and content-recommendation services.
Entity Extraction
April 1, 2022 – June 1, 2023
Led the development of Transformer-based entity extraction models for user authentication and intent understanding, now fully deployed in production across real-time dialog systems. Implemented a token-free Transformer email extraction model that identifies user credentials from ASR transcripts with high precision and sub-second latency, ensuring seamless real-time authentication. Delivered substantial accuracy and latency improvements over baseline models, contributing to more reliable and efficient user verification within conversational products.
Automatic Autism Detection (NSF-Supported)
June 1, 2010 – May 1, 2013
Autism forms a group of severe neuropsychiatric conditions. However, diagnostic procedures to determine whether this problem is present are expensive, require special expertise, and are unreliable. Using technologies that accurately analyze crossmodal affective signals could revolutionize what computers can do for children with Autism in terms of assessing and training their core deficits. The project consists of developing a range of interrelated algorithms that analyze the facial, vocal, and word-content modalities, and automatically determine incongruity of the affective information in these modalities. My role is to develop methods for frontal view search among multiple cameras, facial points tracking, head pose estimation and shape constrained nonrigid point set registration for facial expression analysis.
HairDresser – Automatic Hairstyle Design System based on Face Processing
August 1, 2007 – May 1, 2010
The system can design a suitable hairstyle for the customer by a trained machine which is similar to the experience of large number of hairdresser experts. Firstly, the system can detect and analyze the geometric structure of the head and face by my proposed face processing algorithms. Then the analyzed data can be extracted to be a unique feature. Finally, the feature is input into the trained machine to get a hairstyle that recommend.
Video‐based Watch‐list Face Recognition Surveillance in Shanghai Metro (National Project for 2010 Shanghai Expo)
August 1, 2007 – May 1, 2010
This project is to build a face recognition surveillance system by CCTV for 2010 Shanghai Expo. It can offer a partially accurate means to identify potential terrorists and criminals. In the testing period, the run time of the system is 3 weeks. There are 47 target persons and 952 non-target persons. In addition, uncountable strangers joined into the test implicitly. Every day, each target person passed the region (left image) captured by the system 10 times at least. For the system, only one face image is enrolled for every target and non-target person. However, the probe is real-time surveillance video sequences. I joined to develop the video-based face recognition systems and further improve the system for unconstraint scenarios.
KnowU – Face Recognition System in Video
September 1, 2004 – May 1, 2010
KnowU can be divided into 7 parts: 1. The module of capture: The module programmed based on Directshow is to capture image from video sequences. 2. The module of face and eye detection: Here I adopt Adaboost to detect faces. Template Matching Algorithm is applied to the eye location. 3. The module of face tracking: The algorithm adopted is Mean Shift. 4. The module of preprocessing: Eliminating illumination method is based on Histogram Equalization or Simple Illumination Processing Chain. 5. The module of feature extraction: The module is based on (2D)2PCA or LBP. 6. The module of classification: The module is based on BP with Impulse or Weighted Chi Square Statistic. 7. The module of database management: The module is based on ADO+Access technique.
Breast Cancer Detection based on Microcalcifications
September 1, 2004 – July 1, 2007
This project is to detect breast cancer based on microcalcifications. I did this project during my Master student period. I proposed a corresponding algorithm. The procedure of this algorithm can be divided into three stages: (1) locate the clusters of microcalcifications in mammograms; (2) extract the features of the clusters; (3) classify the features. The result is to justify if the detected clusters are malignant or benign.
Cultural Fit Analysis
The candidate has worked at major tech companies (Google DeepMind, AWS) and led R&D teams, indicating adaptability to fast-paced, innovative environments. The diversity of projects, from personalization and user activity prediction to medical image analysis and surveillance, suggests a broad interest and ability to apply data analysis skills across various domains. The academic background (Ph.D., Master's) further supports a research-oriented and continuous learning mindset. The target role of 'Data Analyst' might be a slight under-match for the candidate's extensive ML/Deep Learning background, potentially indicating a desire for a more focused analytical role or a stepping stone.
Soft Skills & Operational Fit
The candidate's project descriptions highlight leadership in design and implementation, suggesting strong problem-solving and execution skills. Experience as a Team Lead and Research Engineer/Development Manager indicates leadership and project management capabilities. The career break for full-time parenting shows a commitment to personal responsibilities, which can be viewed positively for work-life balance understanding.