Building systems that fuse rigorous economic theory with production-ready machine learning. Currently at Instacart in San Francisco, focusing on marketplace optimization, availability forecasting, and growth experimentation—systems that handle billions in GMV and touch millions of customers.
Dr. Jonathan Louis Gu is an economist and machine learning engineer who builds systems that fuse rigorous economic theory with production-ready machine learning. Currently based in San Francisco and working at Instacart, he focuses on marketplace optimization, availability forecasting, and growth experimentation—systems that handle billions in GMV and touch millions of customers. His work sits at the point where experimental design, causal inference, and large-scale engineering meet, ensuring that complex marketplaces not only function but deliver reliably on what customers expect.
Jonathan's professional approach is defined by breadth and precision. He has designed and deployed causal measurement frameworks that capture subtle, long-term effects without sacrificing decision speed. His economic mechanism work draws on auction theory, multi-agent incentives, and market design, enabling robust policy choices that perform well in dynamic, competitive environments. In machine learning systems engineering, he favors reproducible pipelines, JSON-configured training, and feature store integration—making high-impact models auditable, observable, and low-maintenance.
With a PhD in Economics from UCLA and a BA in Statistics and Economics from UC Berkeley, Jonathan brings an academic depth that is rare in industry. His doctoral work corrected conventional update rules in reinforcement learning to better capture long-run value across states. His research also includes regression discontinuity designs and control-function approaches for education policy, as well as the optimization of bidding strategies in search auctions.
Jonathan's professional philosophy is grounded in adaptability and interdisciplinary thinking, shaped by a multicultural life across Hong Kong, California, and Boston. He believes in "树挪死, 人挪活" ("Tree Move Dead, Man Move Live")—a readiness to change environments, tools, and perspectives in pursuit of better solutions. In his view, the most valuable systems are those that endure with minimal upkeep, because they are designed from the start to work with the grain of human and market behavior.
Frame hypotheses, design trustworthy A/Bs, and translate results into product and policy. Write clear memos with decision ranges and risks.
Availability/forecasting pipelines that are reproducible, auditable, and fast to iterate—so reliability is a habit, not a quarterly project.
Auctions and multi‑agent incentives. Long‑run value beats short‑term hacks; simulate sequences, not snapshots.
Python/SQL/Julia. Evaluation harnesses, drift checks, and low‑maintenance production.
Designed a full‑funnel study of a subscription promotion, modeled day‑of‑week effects, and presented a decision range with risks and guardrails.
Standardized training via JSON configs tied to feature tables; added eval harnesses and alerts so on‑call knew when, why, and how to retrain.
Worked with research groups on search auction mechanisms and bidding strategies. Emphasis on long‑run value, incentive alignment, and policy robustness.
Explores update rules that prioritize long‑term value across states over immediate rewards.
Applied causal designs to student decision‑making and outcomes.
On adaptation across disciplines and cultures—why movement creates life.
Building ML systems with max reliability, speed, and low latency. Customers know when items are in stock with pinpoint accuracy. Translate experiments and models into shipped product decisions. Focus on reliability, measurement quality, and speed of iteration.
A tiny game about fetch and friendship. It's muted by default and keyboard‑accessible.