Invited Talk
VP of Machine Learning
Title: Recent advances and learnings on optimizing the 2-sided marketplace on Upwork
Abstract: Optimizing search ranking on freelancing platforms like Upwork is critical for connecting clients with best-fit talent while driving key business outcomes such as increased hiring rates and revenue growth. Traditional ranking systems often struggle to align with complex user preferences and handle noisy, multi-objective data in dynamic talent marketplaces. To address these challenges, we present a scalable application of Direct Preference Optimization (DPO) tailored for Upwork’s search ranking framework. Our approach leverages advanced DPO techniques, including S-DPO for multi-negative sampling and 𝛽-DPO for adaptive calibration, to align rankings with user intent while dynamically managing data quality variations across multilingual profiles. By fine-tuning Large Language Models (LLMs) with Low-Rank Adaptation (LoRA), we enable efficient real-time ranking updates, ensuring scalability for high-traffic production environments. This work highlights the potential of preference-optimized LLM-based ranking systems to enhance business impact in competitive online talent platforms, particularly for users with sufficient interaction histories.
Invited Talk
Product Data Scientist Lead
Title: Optimization at Waymo: Fleet Scheduling, Positioning and Matching
Abstract: In this talk, I will discuss some of the optimization challenges we face running a ride-hailing fleet of autonomous vehicles at Waymo. The talk will focus on the fleet scheduling problem, where we optimize the launch and return schedule of the cars. We try to optimize the in-service time of cars in a way that matches the variable demand throughout the day, subject to several operational constraints. On the marketplace side, I will explain some of the challenges we face optimizing the position of the open vehicles and matching riders with vehicles
Invited Talk
Applied Science Manager
Title: Optimizing for Tomorrow: Long-Term Consumer Pricing Strategy at Uber Eats
Abstract: Optimizing pricing decisions in a high-frequency, two-sided marketplace like Uber Eats requires navigating complex trade-offs between short-term performance and long-term strategic outcomes. A key objective is to grow consumer lifetime value (LTV)—a function of order frequency, basket size, retention, and profitability. However, direct optimization of LTV is intractable due to delayed feedback loops, sparsity in counterfactual observations, and weak, noisy causal signals from interventions to long-term outcomes. In this talk, we present our technical approach to consumer pricing under a long-term growth objective with profitability constraints. We discuss the design and validation of proxy metrics that serve as short-term surrogates for LTV, chosen based on their sensitivity to product and algorithm changes, alignment with long-term directional goals, and ability to represent temporal behavioral dynamics. We cover our use of causal inference methods, offline validations, and online experimentation for online price deployment.