Allie Cellber

Data Scientist, Machine Learning (Ii -- Iii) @Next Insurance

Palo Alto, CA, US

Owner of the most advanced and impactful series of risk ML models. This set of projects has reduced losses by ~$20M/year from a combination of declining risky policies (2022-present), adjusting users’ exposure (2023-present), and being used as a pricing factor (2024-present)— Modeling: Developed and pioneered xgboost regression models that have predicted loss ratio (20••••24) and now loss amount (2024-present). Specialty techniques used have included building multiple subpopulation-specific models that are trained on the full dataset but weight similar policies more heavily. Increased the key model evaluation metric, gini, from.18 to.30 through these efforts. Went through extensive model evaluation, explanation, and documentation processes to get buy-in from the VPs of insurance product and actuarial teams— System Architecture: Designed and set requirements for multiple engineering squads (as well as ML) to build a new rating capability that supports the quote-time risk differentiation based on loss prediction. Also defined and helped implement data logging and data model requirements— Deployment: Led development of 15 RESTful API endpoints across 3 versions, handling 10k calls/day with success rate and <0.5s latency. The key technologies used include Python, FastAPI, Alembic, PostgreSQL, Poetry, Docker, and Git-lfs— Training Infrastructure: Accelerated ML training by optimizing hyperparameter tuning and bootstrapping using high-CPU SageMaker notebooks. Leveraged S3 and Boto for efficient artifact transfer between environments— Monitoring: Implemented comprehensive monitoring for model inputs/outputs using Arize and Airflow, and created Tableau dashboards for action and model performance monitoring— Business Focus: Personally identified opportunities where this set of projects can be applied to accomplish top business objectives, and collaborated and iterated efficiently to exploit them.• More (cut off by character limit)