Deep learning has gotten considerably easier in the past few years, with open source software, cloud-based infrastructure, and pretrained networks enabling fast and accurate prototyping of AI applications. But what happens when your AI task quickly becomes nontrivial? Brian Dalessandro and Chris Smith discuss Zocdoc’s road to a production-ready, reproducible AI system for insurance card classification and OCR.
Zocdoc is an app that allows patients to find and book doctors online. In additional to facilitating online medical scheduling, Zocdoc makes it easy for patients to check in online and precomplete their medical paperwork. Its insurance card scan product uses an uploaded insurance card to infer the patient’s insurance plan and verify coverage for their medical appointments—a computer vision task that is a natural fit for AI applications.
Early experiments revealed that off-the-shelf models would not drive the needed level of accuracy, especially for the task of localizing key fields and extracting text from images. After months of experimentation and generally “dark art” fiddling, Zocdoc reached its end-to-end accuracy goals using a multitask architecture that combined CNNs and RNNs with attention mechanisms. The next challenge was to reverse engineer its own work and build a system that was fully reproducible and easily extendable to enable automated and iterative improvements.
Brian and Chris offer a general overview of Zocdoc’s really cool, great-for-patients, state-of-the-art classification and OCR system, focusing on general lessons and tips for promoting your AI system from the lab to a fully automated, production-level, self-learning AI pipeline.
Brian d’Alessandro is a Sr Director of data science at Capital One (Financial Services). Brian is also an active professor for NYU’s Center for Data Science graduate degree program. Previously, Brian built and led data science programs for several NYC tech startups, including Zocdoc and Dstillery. A veteran data scientist and leader with over 18 years of experience developing machine learning-driven practices and products, Brian holds several patents and has published dozens of peer-reviewed articles on the subjects of causal inference, large-scale machine learning, and data science ethics. When not doing data science, Brian likes to cook, create adventures with his family, and surf in the frigid north Atlantic waters.
Chris Smith is a senior principal software engineer at online doctor marketplace and booking startup Zocdoc, where he is developing deep learning models to help Zocdoc’s patients understand the complex world of medical insurance. Through machine learning, Chris is helping to bring data-driven products to an industry traditionally resistant to change. Over his career at Zocdoc, he’s built numerous patient-focused products, rebuilt the company’s CI/CD systems, helped migrate Zocdoc from the data center to AWS, and set up the first version of its microservice infrastructure.
©2018, O'Reilly Media, Inc. • (800) 889-8969 or (707) 827-7019 • Monday-Friday 7:30am-5pm PT • All trademarks and registered trademarks appearing on oreilly.com are the property of their respective owners. • email@example.com