Architecting a Real-Time Optimization Platform for Driver Positioning Products
At Lyft, we think a lot about trading off between the immediacy and quality of the response in automated decision-making. On one end, machine learning dominates products that require near-instantaneous feedback such as in fraud and customer support. On the other end, we have complicated workflows that crawl through user graphs to derive weekly macro-level business insights. This talk focuses on the “in-betweens” such as driver positioning and rider-driver matching that require time to aggregate market-level signals before any useful decision can be made. We explore the design principles that we have come to recognize in developing scalable infrastructure that enable fast, iterative, Science-heavy model and product development of real-time optimization workflows.
- Science DevOps is every bit as important as building the Science models
- It's important for Research/Data Scientists to develop models with an understanding of the Eng infra and affect its development
- Similarly, it's important for Eng to work closely with Science to understand the infra needs and not over-index on a specific business application.
Hao Yi is a Research Scientist at Lyft. On the Driver Positioning team, Hao Yi leads the development of the optimization framework and models that power the Personal Power Zones and Hot Spots products that replaces the Driver Prime Time dynamic pricing experience. Previously, Hao Yi combated transaction and driver fraud on the Integrity team. There, he championed new approaches and led ML improvements that helped Lyft achieve best-in-class status in fraud rate within the ridesharing industry. On Support Experience, Hao Yi developed deep learning models for support ticket classification and routing that led to massive reductions in false positive tickets and manual ticket re-routing. Before Lyft, Hao Yi worked on drone traffic management with NASA Ames as a graduate at Stanford.