WeightWatchers

Student Data Scientist | Modeling | Development

WeightWatchers Overview

Weight Watchers, now known as WW International, is a global wellness company that provides weight loss and wellness management services. The company was founded in 1963 and is headquartered in New York City. Its primary program, called the WW program, uses a points-based system to help members make healthier food choices and increase physical activity. The company also offers coaching, workshops, and a mobile app to help members track their progress and stay motivated. With millions of members worldwide, WW International aims to inspire healthy habits for real life and help people lead healthier and happier lives.

My Contributions

I worked as a student data scientist @WW, under the modeling development function for the Data Science Team.
My role covers developing new function of audio transcription, allowing app to capture the food entity for user's audio input and map to the WW's Database.
My responsibilities cover the audio transcription API/ NLP API's development, using open sourced whisper / BERT (DeBERTa_V3 model), and I also participated in designing user interface (using Streamlit) to present our model by embedding it into the internal website.
The coolest thing is that we created a separated version allowing Columbia University Student to use!
‍

Job Duties

Student Data Scientist @WW

Jan 2021 - May 2021

Skills: Python (tensorflow, keras, Huggingface Hub, Whisper, DALLE), SQL, UI Design(Python Streamlit), Deep Learning, NLP, Data Structure

In this internship experience, I participated in developing the beta version of an audio transcription model that was embedded as a function in the user input section. To improve the accuracy of the food-entity-detection task, I created 100,000 pseudo data containing text and POS tags and trained an open-source DeBERTa_V3 model, conducting hyper-parameter tuning processes to achieve 92% accuracy. I also adopted the cosine similarity technique of vector embedding combined with the FAISS model to map detected foods to the database.To provide a user-friendly interface for the product team, I designed a beta version of the user interface via a Streamlit app and deployed it on the internal system for evaluation. Additionally, I created a data warehouse using PostgreSQL to record beta user input and results, using Python Pandas and SQL to test the model's performance on the pseudo-audio input and its variance among food and sentence structures.