William Zhengwei Ma

Senior Data Scientist

I am a Senior Data Scientist for the Office of Data Science, a centralized DS team to promote DS excellence across the enterprise.

I am heavily involved in helping engineers create an enterprise standard for DS model deployment. I have in depth experience with the E2E MLOps cycle - I'm familiar with tools like Airflow/Luigi for ETL, MLflow for model training, Flask apps and Docker containers for packaging, Bamboo and Artifactory for builds, AWS Fargate / Sagemaker for deployment with DataDog and Splunk for monitoring. At the same time I am exploring bringing more innovation solutions like Seldon to solve challenges like inference graphs and AB deployments. In partnership with various business teams and data scientists, the models I am responsible for helping deploy create an estimate of $10m+ in annual value.

I'm currently working on an initiative for claims fraud modeling using Neo4J.

I supported NLP initiatives to develop models to assist in email information extraction using tools like Named Entity Recognition, Question Answer, and Natural Language Inference, based heavily on models from HuggingFace Transformers, as well as looking at various tools for extracting data from forms and documents.

For various computer vision initiatives working with aerial imagery, I am focused on creating a sustainable and efficent pipeline for tile architecture . This includes pipeline design with async calls for fast inference as well as image processing for thousands of addresses. Project worth ~ 1m annually.

I was a manager for co-ops, and led development on deploying a NLP model for email triage (estimated 100k+ annual value).

Graduate Teaching Assistant

I was a TA for Advanced Python for Data Science, a Master's program course at Harvard Extension School for the Data Science Masters Degree. I taught software best practices leveraging open-source tools to aspiring data scientists.

I covered topics like CI/CD, unit testing, task workflows, git, containerization, APIs, skeletons - you can find the syllabus here for more topics - in weekly virtual sections and office hours. I also helped improve our homework grading process with more automated grading via unittests.

Senior Data Scientist

MAPS is a project creating underwriting predictions from physician statements. I leveraged hugging face transformer's question answer fine-tuned BERT models (BERT, BioBert, SciBert) for NLP and pytorch pretrained image models (resnet) with Mask RCNN for computer vision. The entire project is pipelined with Airflow (previously Metaflow) for reproducibility and dockerized.

For OCR, we worked with Tesseract and OpenCV to produce higher-quality OCR, with experiments with Calamari and ABBYY.

I built an end-to-end deep learning entity resolution application. Pipelined with Luigi, it's deployed with Flask in a Docker container with Azure single sign on for security and utilizes Azure pipelines for CI/CD. Based on the Deepmatcher project.

Data Scientist II

I developed a machine learning pipeline (> 100GB of data) with dask, luigi, and dask-ml to model the riskiness of drivers using tree-based xgboost tweedie regression model. It leveraged a Bamboo pipeline for CI/CD for unit testing and deploying sphinx documentation.

I built an end-to-end python ETL pipeline for processing/analyzing telematics data using dask, luigi, and spark. Incorporates geohashed hilbert points for spatial join, density based clustering, and geofences built from webscraping.

I prototyped an implementation of mask RCNN for detecting license plates & pools from satellite imagery, and developed an experimental design framework for testing extractive summaries on diarized Amazon audio transcripts.

Actuarial Analyst

Prior to becoming a data scientist I was an actuarial analyst, focused on Solvency II regulation.