Architected and implemented an AutoML system that powered personalized insurance product recommendations, significantly expanding coverage options while optimizing unit economics:

• Designed a sophisticated feature store with temporal consistency guarantees, handling both first-party application data and third-party data (prescription history, medical records) while ensuring point-in-time correctness

• Built feature engineering pipelines using Apache Beam that maintained semantic groupings of features with immutable definitions and explicit versioning for reproducibility

• Implemented shared feature computation logic between training (ETL pipeline) and production (cache with fallback evaluation) environments using Kafka for event processing and DynamoDB for low-latency serving

• Created a dynamic SQL graph system to manage different feature sets for training, testing, and production scenarios

• Developed automated model training workflows with BigQueryML and XGBoost, including cross-validation, champion/challenger comparisons, and integration with visualization tools (Tableau, R, Google Sheets)

• Built a comprehensive monitoring system that published all model predictions to Kafka streams for real-time analysis in BigQuery, enabling continuous drift detection and performance tracking

This system enabled product teams to rapidly implement personalized user experiences that improved both coverage rates and conversion metrics, resulting in a significant contribution increase in our strategic channels.

Led the development of ML-powered underwriting systems while remaining hands-on, substantially increasing automated decisions and reducing reliance on external partners:

• Designed and implemented ML models that augmented our existing rules engine, building a hybrid decisioning system that captured interactions and subtleties missed by traditional rules

• Created a technical compliance framework for model governance that satisfied both regulatory requirements and reinsurer risk controls while enabling rapid deployment of model updates

• Developed a performance analytics layer in BigQuery that continuously monitored model decisions against actuarial targets, enabling data-driven refinement

• Built an ML-based "fast track" case identification system that prioritized applications with high approval likelihood, reducing time to decision by half

• Designed and implemented an audit workflow system that provided quick feedback on model performance despite the inherently long-term nature of life insurance outcomes

This ML-enhanced underwriting technology more than doubled our automated decision rate without increasing risk exposure and enabled the launch of Ladder's proprietary carrier product.

Led the development of Ladder's core rules engine while writing significant production code:

• Designed and led development of a tree-based underwriting rules engine in Clojure, architecting it to be statically analyzable and serializable for full transparency and auditability

• Created inspection tools that integrated with web viewers, PDFs, and Google Sheets to provide clear visibility into decision logic for underwriters, carriers, and internal teams

• Developed a rules authoring interface that enabled non-technical underwriters to define and update rules without engineering involvement, dramatically improving iteration speed

This rules engine transformed Ladder's underwriting capabilities, creating the foundation for our industry-leading automated decision platform.

Designed and implemented core systems for Ladder's initial underwriting stack:

• Built the instant underwriting system from scratch, including a GraphQL implementation in Clojure and a Relay-inspired frontend in ClojureScript

• Developed an evidence ordering system on Kafka that coordinated collection and evaluation of medical records, prescription histories, and other third-party data

• Created the first version of Ladder's automated decisioning capability which formed the basis for future development

This early technical foundation enabled Ladder's transition from a manual to an increasingly automated underwriting approach.