• Created anomaly detection tool that uses statistical hypothesis testing to alert on out-of-distribution feature vectors

• Tool used by Fixed Income desk to ensure robustness of machine learning models for bonds pricing

• Built out orderflow pipeline and risk check applications for Requests For Quotations (RFQs) on Kafka-like architecture (coded in OCaml)

• Pipeline/apps used by Client-Facing Technology team for low-latency, scalable, and robust price quotes to firm’s clients

• Machine Learning at Berkeley (ML@B) is a club dedicated to fostering a vibrant machine learning community through industry consulting, academic research, and education

• Led team of 10-12 people to run club's internal operations and develop strong club culture

• Organized club events, socials, and multi-day retreats to make ML@B a fun and inclusive environment for its members

• Managed and budgeted $100K+ of club finances

• Served on club executive board, helping make decisions regarding recruitment, club vision, and more

• Worked in team of 5 for our client Vyrill, a video commerce company that helps online businesses increase conversions, sales, and market share with a powerful in-video search engine

• Developed model pipeline that is able to accurately detect the demographics (gender, age, race) of the people in a particular video

• Utilized Google FaceNet to generate facial embeddings, RetinaFace for facial detection and localization, DeepFace to extract demographics, and optimal K-means clustering for duplication person detection

• Used Amazon Ground Truth Plus for labelling and generation of validation set

• Managed team of developers for our client Kopernicus, a company which uses artificial intelligence and big data to deliver actionable business insights, market intelligence, and cost savings for Small-to-Medium Enterprises

• Developed novel technique to extract keywords that identify sub-niches among companies of a given industry

• Generated semantic word embeddings for companies using Bidirectional Encoder Representations from Transformers (BERT) language model, along with K-means and Density-based spatial clustering of applications with noise (DBSCAN) algorithms to cluster together companies within a niche

• Experimented with techniques such as TF-IDF, TextRank, RAKE, KeyBERT (with max sum and maximum marginal relevance similarity tests), as well as algorithms such as DeepDPM (Deep Clustering With an Unknown Number of Clusters)