Hi there!👋🏼

Turning dance competition data into something you can actually explore.

A data pipeline and analytics workflow around International Dance League performances, connecting competition results with song-level metadata to understand performance and music-selection patterns.

• Engineered a Python/Pandas pipeline that validated and normalized 45 verified performances and 68 songs into BigQuery tables
• Designed a 10-table BigQuery warehouse: song and performance dimensions and facts, MusicBrainz enrichment, and derived tables for team genre mix, music strategy, and event trends
• Wrote 4 SQL analytical views using joins, aggregations, and window functions to analyze team scores, song usage, and competition-level music trends
• Enriched 57 of 68 songs (84%) with MusicBrainz release date, genre, and artist metadata
• Shipped 2 BigQuery-connected Tableau dashboards (so far) covering team performance and song-selection patterns

A workspace for teams that spend a lot of time saying “wait, which version?”

A full-stack collaboration platform built around how dance teams actually work, from choreography and rehearsals to videos, corrections, and team communication.

• Designed a multi-team data model with role-based permissions, choreography, sets, rehearsals, and video feedback
• Built timestamped video corrections so dancers can tie feedback to the exact moment in a routine
• Architected a PostgreSQL + Supabase backend with Row Level Security so every team sees only its own data
• Developing the web app in Next.js and TypeScript on a structured relational data model

What if finding a restaurant felt more like asking a friend?

A hybrid restaurant recommender for Santa Barbara, built with two teammates. It combines semantic search, content-based filtering, and collaborative filtering, and surfaces the restaurants that several models agree on.

• My part: built the TF-IDF content-based filtering model over 1,036 restaurants, computing cosine similarity across all restaurant pairs on category, price, and attribute features
• Teammates built the ChromaDB semantic search (all-MiniLM-L6-v2 embeddings, HNSW index) and the ALS collaborative filter trained on ~44,000 Yelp reviews
• The combined system reached 0.783 Precision@5 and 0.833 NDCG@5 in team evaluation, surfacing restaurants that several models agree on

Can a model tell useful comments from spam, even when the spam looks different from anything it has seen before?

An NLP classifier for YouTube comment spam, built to generalize across communities and hold up against obfuscated spam.

• Trained a TF-IDF + Logistic Regression classifier on 1,956 YouTube comments, reaching 0.955 F1 on a held-out test set
• Compared word-level, character-level, and augmented TF-IDF representations to resist obfuscated spam, holding 0.938 average F1 on unseen communities, only 1.7 points below in-distribution
• Deployed the classifier as a Streamlit app for interactive comment classification

A little machine learning, a lot of “what if?”

A classification pipeline that predicts potentially habitable exoplanets, focused on model comparison and how class imbalance affected performance.

• Compared SVM, Random Forest, and Logistic Regression, reaching a 0.78 F1 score on an imbalanced problem
• Validated with cross-validation to confirm performance held beyond a single train/test split
• Investigated class imbalance and multicollinearity to understand their effect on model behavior