David is currently unavailable
David Dai
Verified Expert in Engineering
Machine Learning Developer
Seattle, WA, United States
Toptal member since January 29, 2020
Bio
David has extensive experience in building machine learning and deep learning (DL) solutions at top companies, including Apple, Google, and Facebook, unicorn startups, and academia, as he has a PhD from Carnegie Mellon U. He holds multiple patents in DL-based medical imaging tech and large-scale AI systems. David has grown an AI team to 60+ as the director and tech lead.
Portfolio
Apple
Tools, Core ML, Git, PyTorch, TensorFlow, Artificial Intelligence (AI)...
Petuum
Protocol Buffers, Kubernetes, Docker, GraphQL, C++, Spark, TensorFlow...
Facebook
Tools, Facebook, Vowpal Wabbit, C++, Machine Learning
Experience
- Machine Learning - 10 years
- Distributed Systems - 8 years
- Deep Learning - 6 years
- TensorFlow - 4 years
- Natural Language Processing (NLP) - 4 years
- Computer Vision - 3 years
- PyTorch - 2 years
- Generative Pre-trained Transformers (GPT) - 1 year
Preferred Environment
C++, Python, Scikit-learn, PyTorch, TensorFlow, Django, ChatGPT, OpenAI GPT-3 API, Chatbots
The most amazing...
...project I've created was best-in-class DL methods for a chest X-ray to assist radiologists in diagnosis, resulting in the first contract for the company.
Work Experience
Staff Machine Learning Engineer
2019 - 2022
Apple
- Deployed high-performance deep-learning models to run on-device to improve inference speed and user privacy.
- Led the design of CoreML, a deep-learning framework for on device and cloud for first- and third-party developers at Apple.
- Served as a tech lead and helped grow the team from three to ten engineers and researchers.
Technologies: Tools, Core ML, Git, PyTorch, TensorFlow, Artificial Intelligence (AI), Machine Learning
Senior Director of Engineering
2015 - 2019
Petuum
- Worked as part of the founding team and helped to grow the engineering team from less than ten in 2016 to over 60 full-time engineers in the second quarter of 2018 (2Q18).
- Influenced, as a domain expert in machine learning, a trench worker, and a manager, the product design and delivered quarterly internal platform product milestones from 3Q17-1Q18.
- Led the medical imaging team of four and personally developed best-in-class deep learning methods for chest X-rays to assist radiologists in diagnosis. The project led to the first contract for the company.
- Served as the tech lead and managed the back-end design of Petuum's AI platform that combined all stages of machine learning, including data cleaning, feature engineering, data and model visualization, deployment, and monitoring.
Technologies: Protocol Buffers, Kubernetes, Docker, GraphQL, C++, Spark, TensorFlow, Machine Learning
Software Engineer Intern
2015 - 2015
- Contributed more than 10,000 lines of C++ code during my three-month internship and helped initiate a large-scale distributed training system for Facebook newsfeed and ad algorithms.
- Developed a distributed large-scale logistic regression using LBFGS solver and parameter server-based synchronization.
- Benchmarked the implementation against Facebook's internal system and Vowpal Wabbit. Showed that the implementation achieves high system throughputs and produces comparable to better models.
Technologies: Tools, Facebook, Vowpal Wabbit, C++, Machine Learning
Software Engineering Intern
2013 - 2013
- Optimized one of the world's largest machine learning systems – Google's ads algorithms for click-through-rate prediction.
- Developed a large-scale hyperparameter tuning framework to enable Google's ads training system to discover the best hyperparameter settings for ML models. The system is used by dozens of teams and later published as Google Vizier.
- Built the frontend to the hyperparameter tuning engine.
Technologies: Tools, Google, Protocol Buffers, C++, Machine Learning
Software Developer Intern
2012 - 2012
- Implemented a number of background tasks in the payment back end.
- Improved access control for internal payment handling.
- Developed a dynamic terminal folder navigator to replace the "CD" in Linux in a hackathon.
Technologies: Spring, Python, SQL, Java
Experience
Chest X-ray Organ Segmentation Through Adversarial Deep Neural Networks
MOTIVATION
A medical imaging device that a client needed to differentiate their products with AI.
BRAINSTORMING
I collaborated closely with radiologists to identify tasks that can be automated by AI in the near term. While most AI apps in medical imaging aren't interpretable, e.g., black box classification of diseases, the solution needed to be interpretable to build trust with doctors. We chose chest X-ray segmentation, which can be visualized and directly leads to clinical indicators.
DEVELOPMENT
Existing methods use traditional computer vision pipelines, which are very brittle and inaccurate. I found only a minimal dataset with pixel labels (<300 samples), which was challenging for data-hungry neural networks. After intensive studies, I successfully reduced the ImageNet model by 500 times to adapt to a small training dataset and achieved human accuracy for the first time.
DEPLOYMENT AND IMPACT
I led a team of four engineers to produce the model for the client, including building REST API endpoints on public clouds designated by the client for security and latency considerations. I trained the team to fine-tune the model on client data. The developed solution landed the first contract for our company.
A medical imaging device that a client needed to differentiate their products with AI.
BRAINSTORMING
I collaborated closely with radiologists to identify tasks that can be automated by AI in the near term. While most AI apps in medical imaging aren't interpretable, e.g., black box classification of diseases, the solution needed to be interpretable to build trust with doctors. We chose chest X-ray segmentation, which can be visualized and directly leads to clinical indicators.
DEVELOPMENT
Existing methods use traditional computer vision pipelines, which are very brittle and inaccurate. I found only a minimal dataset with pixel labels (<300 samples), which was challenging for data-hungry neural networks. After intensive studies, I successfully reduced the ImageNet model by 500 times to adapt to a small training dataset and achieved human accuracy for the first time.
DEPLOYMENT AND IMPACT
I led a team of four engineers to produce the model for the client, including building REST API endpoints on public clouds designated by the client for security and latency considerations. I trained the team to fine-tune the model on client data. The developed solution landed the first contract for our company.
Wayfinder Vizier | Rich Question Answering via Neural Semantic Search
Wayfinder Vizier enables rich question answering over any text, including customer reviews, social media, articles, and FAQs, and brings contextual information retrieval beyond keyword-based search.
For example, Wayfinder Vizier answers the following questions about Chihuly Garden and Glass, a tourist destination, based on TripAdvisor out-of-box and without any tuning user reviews. The answers, i.e., excerpts from user reviews in this case, are both diverse and relevant.
Q: What's the best time to visit the museum?
• A sunny day is the best time to visit.
• I think the garden at night is the best time to view it.
• The best time to visit is just before the sunset.
• My family of three made our reservations for 7:00 PM. (They request not only the date of your visit, but you must select a time that you will visit.) It was the perfect time to visit.
• Summer is not the best time to visit.
Q: Is it a good idea to bring kids?
• Not a place to bring young kids.
• It may not be worth it to bring the little ones.
• Is a good attraction for families.
• What a great place to visit with kids.
• Not a place for kids.
Wayfinder Vizier was selected as one of 10 Emerging AI Pioneers to present at O'Reilly AI Conference 2019.
For example, Wayfinder Vizier answers the following questions about Chihuly Garden and Glass, a tourist destination, based on TripAdvisor out-of-box and without any tuning user reviews. The answers, i.e., excerpts from user reviews in this case, are both diverse and relevant.
Q: What's the best time to visit the museum?
• A sunny day is the best time to visit.
• I think the garden at night is the best time to view it.
• The best time to visit is just before the sunset.
• My family of three made our reservations for 7:00 PM. (They request not only the date of your visit, but you must select a time that you will visit.) It was the perfect time to visit.
• Summer is not the best time to visit.
Q: Is it a good idea to bring kids?
• Not a place to bring young kids.
• It may not be worth it to bring the little ones.
• Is a good attraction for families.
• What a great place to visit with kids.
• Not a place for kids.
Wayfinder Vizier was selected as one of 10 Emerging AI Pioneers to present at O'Reilly AI Conference 2019.
Education
2012 - 2018
PhD in Machine Learning
Carnegie Mellon University - Pittsburgh, PA, USA
2010 - 2012
Bachelor of Science Degree in Computer Science
Caltech | California Institute of Technology - Pasadena, CA, USA
2007 - 2010
Bachelor of Arts Degree in Physics and Mathematics
Wesleyan University - Middletown, CT, USA
Skills
Libraries/APIs
TensorFlow, PyTorch, Scikit-learn
Tools
ChatGPT, Git
Languages
Python, C++, GraphQL, SQL, Java
Platforms
Linux, Docker, Kubernetes
Frameworks
Core ML, Spark, Spring, Django
Storage
MySQL, RocksDB, LevelDB
Other
Machine Learning, Deep Learning, Distributed Systems, Deep Neural Networks (DNNs), Artificial Intelligence (AI), Computer Vision, Natural Language Understanding (NLU), Natural Language Processing (NLP), OpenAI GPT-3 API, Chatbots, Tools, Protocol Buffers, Vowpal Wabbit, Facebook, Google, Generative Pre-trained Transformers (GPT)
Collaboration That Works
How to Work with Toptal
Toptal matches you directly with global industry experts from our network in hours—not weeks or months.
1
Share your needs
Discuss your requirements and refine your scope in a call with a Toptal domain expert.
2
Choose your talent
Get a short list of expertly matched talent within 24 hours to review, interview, and choose from.
3
Start your risk-free talent trial
Work with your chosen talent on a trial basis for up to two weeks. Pay only if you decide to hire them.
Top talent is in high demand.
Start hiring