Yuhe Zhang
ML EngineerAI is analyzing your overall score…
Identifying your key strengths…
Evaluating your skill match against the job requirements…
Assessing your cultural and operational fit
AI is analyzing your overall score…
Identifying your key strengths…
Evaluating your skill match against the job requirements…
Assessing your cultural and operational fit
Senior Deep Learning Algorithm Engineer at Nvidia
AI researcher and engineer with 8+ years of experience building production-grade machine learning systems for computational photography and creative tools. Currently a Senior Algorithm Engineer on the NeMo training framework team at NVIDIA, focused on large-scale AI training systems and GPU-efficient machine learning infrastructure. Previously, I led the AI team at Polarr (now part of Pixieset), where I worked on ML-powered photo editing systems used by professional photographers and partners including Samsung and OPPO. My work spanned applied computer vision research, model optimization, and deploying real-time AI systems across mobile, web, and on-device platforms. Ph.D. in Computational Physics.
Penn State University
Doctor of Philosophy (Ph.D.), Computational Physics
January 1, 2011 – January 1, 2017
Zhejiang University
Bachelor’s Degree, Physics
January 1, 2007 – January 1, 2011
NVIDIA
Senior Deep Learning Algorithm Engineer
May 1, 2026 – Present
Remote
Pixieset
Staff AI Research Engineer
May 1, 2025 – Present
Remote
Polarr
Head of A.I. Engineering (Staff Research Engineer)
January 1, 2019 – May 1, 2025
San Francisco Bay Area
Polarr
Machine Learning Developer
August 1, 2017 – January 1, 2019
San Francisco Bay Area
Penn State University
Research Assistant
August 1, 2014 – June 1, 2017
Penn State University
Teaching Assistant
June 1, 2012 – August 1, 2014
Triton-Augment: GPU Kernel Fusion for High-Performance Image Augmentation
October 1, 2025 – Present
Developed an open-source PyTorch library leveraging OpenAI Triton to achieve extreme performance gains in deep learning data augmentation. Designed a fused GPU kernel architecture that eliminates the 'Global Memory Tax' of sequential transforms (Crop, Jitter, Normalize). Achieved 5x–12x speedup over modern Torchvision implementations on high-resolution image data. Demonstrates expertise in GPU, low-level optimization, and custom kernel development.
Instant ID Photo Web App
April 1, 2024 – May 1, 2024
End-to-end launch of a free, independent web product for instant passport and visa photo compliance. This project showcases full-stack AI deployment and product ownership with zero server overhead. * Client-Side AI: Deployed core computer vision logic—including face landmark detection and background removal—to run entirely in the user's browser. This ensures maximum user privacy and lightning-fast processing speeds. * Automated Compliance: The application automatically handles the entire pipeline: image straightening, background removal, precise cropping, and generating print-ready collages to meet official standards. * Solo Product Launch: Demonstrated the ability to solve a real-world problem, build the full web stack, deploy, and scale the application independently.
Recognition of Sequences of Letters from Images (Python, Tensorflow)
August 1, 2016 – January 1, 2017
1. Built a single-letter recognition model using convolutional neural networks and obtained an accuracy of 97% on the public dataset notMNIST. 2. Implemented a deep learning model that operates directly on the image pixels to identify multi-character sequences with varying lengths (based on Goodfellow et al. 2013). This approach avoids the traditional separated steps of localization, segmentation and recognition, and achieves an accuracy of 90% on synthetic sequences of letters generated from notMNIST.
Urban Blight Prediction (Python, Scikit Learn)
June 1, 2016 – November 1, 2016
Urban blight refers to the deterioration and decay of buildings and older areas of large cities, due to neglect, crime, or lack of economic support. The city planners are actively trying to predict which properties are likely to become officially classified as blighted ahead of time. This can help the city take preventative actions. In this project, I work with real data from the Open Data Portal of Detroit to help urban planners build a blight prediction model. This is a real-world problem: the data is not perfectly clean, the questions are not perfectly unambiguous. There are two major things that I did in order to build a good blight prediction model using Scikit Learn: 1. Created a list of buildings from all the geo-located incidents collected through the open data portal; 2. Improved the accuracy of the model by extracting a richer set of features and selecting the best model from SVM, Decision Tree, Random Forests by cross validation. A ipython notebook is available in the following link including the entire process of data acquisition, preprocess, modeling and analysis: https://github.com/seedlingfl/Blight-Fight
Twitter Sentiment Analysis (Python, Twitter API)
April 1, 2016 – August 1, 2016
1. Estimated the public’s perception (the sentiment) of a particular term of phrase. 2. Analyzed the relationship between location and mood based on the live stream of twitter data and found that Indiana is the happiest state.
Machine Learning Projects
February 1, 2016 – November 1, 2016
A series of projects to implement various machine learning algorithms from scratch in MATLAB including: 1. Linear Regression: predict housing prices in Portland, Oregon using Gradient Descent algorithm with feature normalization. 2. Logistic Regression: implement regularized logistic regression to predict whether microchips from a fabrication plant passes quality assurance. Apply feature mapping to generate polynomial terms in order to obtain an accurate non-linear boundary. 3. Multi-class Classification and Neural Network: implement a one-vs-all logistic regression model and the feedforward propagation part of a nerual network to recognize hand-written digits. 4. Neural Networks Learning: implement the backpropagation algorithm for neural networks and apply it to the task of hand-written digit recognition. 5. Regularized Linear Regression and Bias v.s. Variance: implement regularized linear regression and generate the learning curves to study models with different bias-variance properties. 6. Support Vector Machines: use support vector machines (SVMs) to build a email spam classifier. 7. K-means Clustering and Principal Component Analysis: (a) implement the K-means clustering algorithm and apply it to compress an image. (b) implement principal component analysis to find a low-dimensional representation of face images. 8. Anomaly Detection and Recommender Systems: (a) implement the anomaly detection algorithm and apply it to detect failing servers on a network. (b) implement collaborative filtering to build a recommender system for movies. All the codes can be found at: https://github.com/seedlingfl/Machine-Learning-mini-Projects
Development of Visualized Numerical Simulation using Java
September 1, 2013 – December 1, 2013
Developed numerical simulation for a variety of physical problems (such as particle motion, optical diffraction, thermalization, percolation) with visualization using Java. As a teaching assistant at that time, I showed some of the simulation demos to the students, which improved their course performance and stimulated their interest for advanced topics in physics.
Machine Learning by Stanford University on Coursera
Stanford University on Coursera
June 24, 2026 – Present
Cultural Fit Analysis
The candidate's project portfolio demonstrates a strong passion for machine learning and AI, with diverse applications from low-level GPU optimization to full-stack AI product launches. Their experience in both academic research and fast-paced startup environments, including a successful acquisition, suggests a versatile and growth-oriented mindset. The open-source contributions and personal projects indicate initiative and a continuous learning attitude, which aligns well with innovative tech cultures.
Soft Skills & Operational Fit
The candidate's experience as a 'sole AI Research & Development Individual Contributor' and leading small teams indicates strong autonomy, problem-solving, and leadership potential. Their ability to drive projects from concept to production in an evolving startup environment suggests adaptability and a results-oriented mindset. The teaching assistant role also implies good communication and mentorship skills.