Claire Hester

Face Mask Detection

The worldwide COVID-19 pandemic has led to the widespread wearing of face masks to protect ourselves and others from spreading the disease.  For my Data Science Immersive capstone project, I created an object detection model that can identify whether a person is wearing a mask correctly, incorrectly, or not at all. To solve this problem, I took a two-step approach: first, use a face detector model to locate all of the faces in an image or video. Second, build a face mask classifier using MobileNetV2.  

For face detection, I used two different models. MTCNN  had a higher degree of accuracy and was able to detect a higher number of faces.  I also used an OpenCV built-in DNN. Though this model had a lower accuracy, the frames per second was 4 times faster than the MTCNN model.  The face mask classifier was built on a MobileNetV2 base with a custom head layer. I achieved a 94% accuracy after 25 epochs, with a loss of 0.19. The "With Mask" category had the most samples for training and had a recall of 99%.
To learn more about this project, click here.

California Wildfires: Measuring the Economic Impact

Catastrophic wildfires in California have sweeping impacts: environmental costs, poor air quality, carbon emissions, homes lost, large numbers of people displaced, and suppression costs, to name a few. Our goal is to build a convincing argument for the redirection of funds to catastrophic wildfire prevention. We focused on three factors for prediction: acres burned, fire suppression costs, and structures destroyed over the next 10 years.

Using an ensemble time series model we predicted that nearly 13 million acres will burn, which will lead to 61,000 structures destroyed and fire suppression costs of almost $40 billion. As it currently stands, there are 10 million acres of forest that are considered "high risk." At a forest restoration cost of $2350 per acre, we believe that if funds were allocated to forest restoration we would see an overall savings of $16 billion in public funds over the next 10 years. To learn more about this project, click here.

Reddit Classifier using Natural Language Processing

Reddit calls itself the front page of the internet, and one very popular use of Reddit is asking questions. The goal is to build a classifier that can determine which subreddit a post is from, given the text of the post. I used the Pushshift API to scrape 60,000 posts from the Legal Advice and No Stupid Questions subreddits. Data was cleaned to address removed posts, missing values, and duplicates, as well as special characters and text features.

I used Vader sentiment analysis to explore key differences, as well as word frequency by subreddit. CountVectorizer was used to turn all words into numerical values. I tested out several different types of classification models and utilized GridSearch and pipelines to hypertune parameters. Logistic regression performed the most strongly, with a 94% test accuracy score. To learn more about this project, click here.

Data Analyst

Use Python and SQL to build complex client-facing reports. Built a process for extracting text from scanned PDF documents, thereby replacing a manual data entry process. Took initiative to learn laTeX document creation and taught a team workshop, unlocking the ability to automatically build and format complex PDF reports.

Data Science Fellow

480-hour immersive program applying data collection and cleaning, analysis, modeling, data visualization, and machine/deep learning techniques to solve real-world data problems. Used Natural Language Processing to analyze and classify 60,000 Reddit posts. Completed client project forecasting the cost of California wildfires using time series analysis and Tableau. Capstone project utilizing computer vision and neural nets to build a face mask detection model.

Associate Planner

Responsible for inventory allocation and management for 45+ stores aimed to maximize sales. Managed inventory accuracy for all retail stores and warehouses across our ERP and POS systems. Led initiative to ensure data integrity and improve system accuracy through investigation, testing, and working directly with the Director of Finance.

North America Wholesale Account Manager & Corporate Sales Representative

Managed 40 accounts including Zappos, Backcountry.com, and The Sports Basement. Created and executed growth and revenue strategies alongside the Head of Global Sales. Developed merchandising and marketing plans aligned with company initiatives. Streamlined and executed logistics for biannual Global Sales Meeting and seasonal trade shows. Managed development of top growth accounts through outreach, research, and custom tailored product recommendations.

General Assembly

Colorado College