Masooma Ali

Designed and deployed an algorithm to determine the semantic and lexical alignment between an investor pitch and a company's marketing material. The algorithm combined traditional ML techniques with a deep learning embedding model to generate insights on the amount of time in the pitch aligned to the marketing material (i.e., on-topic time). I highlighted the sentences on a displayed transcript aligned with the marketing material.

Developed a multi-modal deep learning model to identify the active speaker in a video. The input to the model was prepared by stacking ten video frames per second and providing the tracks for detected objects. The model I created combined these images with audio to predict the bounding box around the current speaker if the speaker was detected on screen.

Sesh was looking to adapt its technology to the parenting domain and needed to identify a core set of pressing problems for parents. I scraped over 500,000 posts from parenting forums in North America. I analyzed them with NLP techniques to create a list of topics parents discussed and defined sub-categories to identify parenting issues. To model the topics, I used word embeddings and clustering techniques (UMAP) combined with lexical chaining for sub-categorization. This project had a big data visualization component as I had to present my findings to the product team in an easily consumable format. This work directly impacted the product roadmap for Sesh's offering for parents.