Chris Dobson

I designed and built a production recommendation platform for a top-20 U.S. insurance provider, delivering upsell and cross-sell opportunities to thousands of call center representatives across millions of customer policies. The system generated recommendations in under 800 milliseconds during live interactions and was credited with approximately $80 million in additional premium revenue.

I architected the platform from scratch using a Node.js/Express API on AWS ECS with a React-based agent interface delivered via static web apps. I implemented a data pipeline where batch ML scores from a Bayesian model streamed through Kinesis into Aurora, with Redis caching to meet real-time latency requirements. The system integrates with multiple enterprise policy APIs to dynamically validate eligibility.

My work included addressing major performance challenges caused by 30–50 downstream API calls per request by introducing intelligent caching and predictive pre-warming based on renewal timelines. I also collaborated with data scientists to operationalize models and worked with stakeholders and agents during beta testing to refine workflows, making the platform a core component of customer service touchpoints.

I designed and implemented a real-time real estate data platform using an event-sourced architecture to support large-scale ingestion, search, and analytics. Property and listing events stream through Apache Pulsar, providing durable pipelines and reliable synchronization across services. I built a custom Go service that materializes data into an in-memory DuckDB engine with geospatial indexing, enabling sub-second map searches and complex filtering across large datasets with high concurrency.

I architected an image ingestion system processing over 20TB of listing photos, including normalization, deduplication, metadata extraction, and vector embedding generation. Images are packaged into ML-ready datasets stored in object storage, supporting downstream computer vision and model training workflows.

My work included designing the foundation for a machine learning-driven property pricing model, combining structured attributes, historical transactions, and image-derived features. The system supports continuous data collection and retraining through event pipelines, enabling automated valuation insights and pricing recommendations.