Prasana Y Doshi

About

I spend most of my time on the parts of machine learning that decide whether a model ever gets used. Data arriving in formats nobody agreed on. Labels that turn out to be judgment calls. The serving path between a trained model and the person who has to act on it. Most of that work has happened at Munich Re's specialty insurance arm, though very little of it is specific to insurance. I trained a gradient boosted model over 3M+ rows of tabular data and brought MAPE from 19% to 7%, cutting the feature set from 47 to 23 with SHAP. I built a general ledger pipeline on top of our IFRS 17 accounting data so finance owners can pull their own headcount numbers, budget against actuals, and expense forecasts without filing a request with us (more projects on my website). I am in a dual degree, CS at Waterloo and business at Laurier, graduating late 2027, and I start at Jobber as a Product Data Scientist in September. I am most interested in applied ML that ships: models with a serving story, retrieval systems with an evaluation set, and problems where the messy part is the data rather than the architecture. Work and thoughts at prasanaydoshi.com Contact at prasanaydoshi@gmail.com

Experience

Education