My work
Product strategy, analytics and machine learning, and design thinking
Problem: High user drop-off during onboarding. Strategy: Ran structured customer discovery across 4 B2C AI product lines. Solution: Prototyped functional workflows for a 3B-parameter domain LLM. Impact: Validated Product-Market Fit and delivered data-backed pricing recommendations.
Problem: Medicare fraud happens per provider, but the data arrives as 558K individual claims. Strategy: Engineered provider-level features and handled a 9% fraud rate with class weighting. Solution: Random Forest scored on PR-AUC, plus a Tableau dashboard. Impact: PR-AUC of 0.706, about 7.6x better than chance.
Problem: A specialty oncology drug maker needed territory sales and account-level rebates after quarter close. Strategy: Cleaned ~10.9K sales rows, mapped accounts to systems and territories, and applied GPO contract tiers. Solution: An Excel rebate engine. Impact: Sized $7.25M in quarterly rebates against $14.6M profit.
Problem: Identifying UX friction and market gaps in mature, high-complexity products. Strategy: Applied structured PM teardown frameworks. Solution: Delivered a 23-slide Microsoft Teams breakdown. Impact: Set a North Star metric and wireframed an AI notification fix.
Problem: A star rating says whether patients liked a drug, not why. Strategy: Read 215K Drugs.com reviews with NLP and linked side effects to ratings and drop-off. Solution: Adjusted VADER sentiment and side-effect detection with negation handling. Impact: Found 35 hidden side-effect signals in well-rated drugs.
Problem: MBA students must upskill, but the time is scattered across 40- to 90-minute gaps nobody plans. Strategy: Design Thinking, from user interviews to a testable concept. Solution: A prototype that ghost-blocks calendar gaps with the next certification module. Impact: A live demo to test the interaction.
Problem: Real data rarely arrives as one clean CSV. Strategy: Rebuilt 161K Drugs.com reviews as a small data warehouse. Solution: Extract, clean and model into a star schema in SQLite, then answer business questions in SQL. Impact: Surfaced conditions whose reviews read far worse than their star ratings.