I have a PhD in economics, search-auction research background, and production ads bidding experience from Instacart. The work I keep coming back to is simple to state and hard to operate: convert noisy marketplace signals into decisions that advertisers, customers, and the business can trust.
At Instacart, I designed and built the company’s first optimized ads automated bidding system. Before it, advertisers manually set bids across product-keyword combinations. The system changed the interface to: tell us your weekly budget and objective, and we will bid intelligently on your behalf. The system controlled over 60% of advertiser spend and indirectly influenced roughly 20% of ads revenue.
The core idea was simple: optimal spend should equalize marginal ROAS across an advertiser’s portfolio, subject to budget and marketplace constraints. That turned the system into large-scale ROAS estimation plus advertiser-level target/control selection.
The production version was a practical decision loop: estimate returns, choose advertiser-level target/control policies, simulate bids, generate bids, respect budgets, observe outcomes, and adjust. It ran through Airflow DAGs, used Snowflake heavily, and included bid-simulation logic in Snowflake UDFs.
The design survived years of iteration: product-keyword ROAS, new advertiser objectives, richer model infrastructure, and additional ad formats.
Before Instacart, I worked with Prof. Susan Athey at Microsoft Research on Bing search-auction counterfactuals and revenue effects under alternative ranking rules.
At Instacart, related work included causal coupon measurement, in-app marketing decisioning, synthetic treatment effects for targeting, and auction/marketplace patent work.
I treat automated bidding as a control system, not just a model. The model score only matters if the system can safely turn it into a decision with calibration, pacing, simulation, monitoring, rollback, and debugging.
That same operating style carries into availability prediction, fulfillment decisioning, and real-time model-serving paths.
OpenClawBrain is also the way I think about LLM systems generally: let the model propose semantic meaning, but make code enforce scope, trust boundaries, retrieval limits, and evidence.
The goal is not a chatbot. The goal is a better decision loop.
For ads, that means using frontier models where language and semantic understanding create leverage: advertiser intent, creative understanding, product/category semantics, campaign diagnosis, explainability, policy/safety, and support workflows — then wrapping those outputs in measurable bidding, ranking, pacing, and causal-evaluation systems.
US 12,511,677 B2, “Automated Policy Function Adjustment Using Reinforcement Learning Algorithm”, issued Dec. 30, 2025, covers reinforcement-learning automated bidding for content campaigns. Assignee: Maplebear Inc.; inventors include Jonathan Gu.
Additional portfolio spans double-wide ad auctions, synthetic treatment effects for targeting, dynamic offer targeting, retailer classification using sales data and LLMs, fulfillment optimization, replacement agents, and item display decisions.
I did not just tune a model. I turned manual advertiser bidding into a production decision system.
Download my resume or return to jonathangu.com.