I have a PhD in economics and production ads bidding experience from Instacart. The work is simple to state and hard to run: turn noisy marketplace signals into bidding and targeting decisions that advertisers, customers, and the business can trust.
At Instacart I built the first optimized ads bidding system from scratch. Before it, advertisers set manual bids for each product and keyword, with only bulk-edit tools around the edges. The new system let them set a weekly budget and a goal. The platform then bid automatically to maximize sales. It now handles roughly 70% of Sponsored Products spend, about $700M annualized.
The premise was ROAS equalization. If one product earns more per ad dollar than another, move spend toward it until returns converge. That split the work in two: estimate ROAS across products and keywords, then choose the right target ROAS for each advertiser.
In production it became a loop: estimate returns, choose target ROAS, pace budgets, simulate bids, write bids, observe outcomes, and adjust. Airflow ran the DAGs. Snowflake held the data. UDFs handled bid simulation close to the data.
The original design survived after I moved on. Later teams extended it to product-and-keyword ROAS, objectives beyond total sales, richer models, and more ad formats.
Before Instacart, I worked with Prof. Susan Athey at Microsoft Research on Bing search-auction counterfactuals and revenue under alternative ranking rules.
At Instacart, related work included causal coupon measurement, in-app marketing decisions, synthetic treatment effects, and the MVP for uplift-based growth targeting: using causal experiment outcomes to rank users by expected incremental treatment effect rather than raw conversion propensity.
I treat automated bidding as a control system, not a model score. A score matters only if the system can turn it into a safe decision with pacing, simulation, monitoring, rollback, and debugging.
I use the same style in availability prediction, fulfillment decisions, and real-time model serving.
OpenClawBrain reflects how I think about LLM systems: let the model suggest meaning; make code enforce scope, limits, and evidence.
The goal is not a chatbot or a score. The goal is a better decision loop.
For ads, frontier models help where language matters: advertiser intent, creative understanding, product meaning, diagnosis, explainability, policy, and support. The hard part is still the wrapper: bidding, ranking, pacing, measurement, and rollback.
US 12,511,677 B2, “Automated Policy Function Adjustment Using Reinforcement Learning Algorithm”, issued Dec. 30, 2025, covers reinforcement-learning control for automated bidding in content campaigns. Assignee: Maplebear Inc.; inventors include Jonathan Gu.
Additional portfolio spans double-wide ad auctions, causal targeting, dynamic offers, retailer classification with LLMs, fulfillment optimization, replacement agents, and item display decisions.
I did not just tune a model. I built the system that made the bids.
Download my resume or return to jonathangu.com.