Jonathan Gu

I build decision systems for real-world uncertainty.

At Instacart, I build systems where models have to become decisions: optimized ads bidding, causal marketing measurement, availability prediction, fulfillment decisioning, and production ML infrastructure. My automated bidding work controlled over 60% of advertiser spend and indirectly influenced roughly 20% of ads revenue.

The throughline is simple: take noisy signals, decide what matters, and keep the system honest once real users and real money touch it. My first-principles ads bidding work is reflected in issued patent US 12,511,677 B2.

Outside work, I build Bountiful Garden for neighborhood harvest sharing, OpenClawBrain for local agent memory, and Project Pelican for autonomous trading research.

Current Role

ML Engineer & Economist at Instacart; built the first optimized ads automated bidding system, controlling over 60% of advertiser spend.

Training

PhD Economics, UCLA. BA Statistics & Economics, UC Berkeley. Still using those causal and decision-system instincts in production ML.

Current Builds

Bountiful Garden, OpenClawBrain, and Project Pelican — products about trust, memory, and decision-making under uncertainty.

Ads Systems Instacart Work Selected Projects Resume
At Instacart

Ads, availability, and marketplace decisioning at Instacart.

My Instacart work spans ads bidding, causal marketing measurement, availability prediction, fulfillment decisioning, and production ML systems. The shared problem is always the same: noisy signals are not enough. The system has to decide what to do, operate reliably, and stay measurable once real users touch it.

What I actually do

This is end-to-end production ML and decision-systems work: first-principles problem framing, data and features, model/control design, batch and real-time serving, experiments, rollout, monitoring, debugging, and migration work when old paths need to be retired.

  • Designed and built Instacart's first optimized ads automated bidding system, moving advertisers from manual product-keyword bids to budget/objective-based automated bidding controlling over 60% of advertiser spend.
  • Framed bidding around marginal ROAS equalization, then turned it into production Airflow/Snowflake infrastructure with ROAS estimation, target/control selection, bid simulation, and budget-aware bidding loops.
  • Developed causal marketing measurement and in-app decisioning work where observed conversions were not clean labels and overlapping interventions had to be handled carefully.
  • Built availability and fulfillment systems across noisy inventory signals, real-time LightGBM serving, calibration, rollout, monitoring, and operational debugging.

Optimized ads bidding

Budget/objective-based campaign optimization, ROAS estimation, bid simulation, and production bidding loops for a real ads marketplace.

Causal and marketplace measurement

Coupon and in-app marketing decision systems where incrementality, selection bias, and overlapping interventions matter.

Availability and fulfillment ML

Models and serving paths for product-store availability, real-time decisions, calibration, forecasting, and operational signal integration.

Selected Projects

Three projects I keep coming back to.

These are the clearest examples of the problems I like: local coordination, better memory for agents, and systems that keep making decisions on their own.

Bountiful Garden neighborhood harvest-sharing interface
Live product Neighborhood software

Bountiful Garden

Bountiful Garden helps neighbors share backyard produce that would otherwise go to waste. The goal is not another community feed. It is to make a simple local exchange actually happen.

It is the most founder-shaped thing I'm building: product design, trust surfaces, and marketplace mechanics for ordinary neighborhood life.

Marketplace Community Full-stack product
Visit Bountiful Garden
Live product Agent memory v0.2.16

OpenClawBrain

OpenClawBrain is evidence, not vibes, for agent memory: it helps an agent understand that “close it out” means proof, remember durable corrections, and stop asking you to repeat the same repo convention twice.

It keeps memory local, retrieves only a small useful slice, and shows its work. Ollama proposes; code disposes. The current public release is 0.2.16, with aggressive audited capture, multi-agent scoped activation, strict scoped storage, recall-rule safety, and bounded SQLite/FTS retrieval verified live.

See OpenClawBrain → How it works Verify it
Project Pelican autonomous options trading research
Private R&D Autonomous trading

Project Pelican

Project Pelican is my private research system for autonomous options trading. It combines data, forecasting, risk controls, and execution into one monitored loop.

It is still internal, but it reflects the same recurring interest as the rest of my work: systems that keep making measured decisions when uncertainty is real.

Quant finance ML pipelines Autonomous systems
Pelican Overview
Patents

Applied ML systems with patent-backed depth.

My Instacart work spans automated bidding, marketplace decisioning, causal targeting, retailer intelligence, and fulfillment optimization — including issued patents and active disclosures.

US 12,511,677 B2 — Automated Policy Function Adjustment Using Reinforcement Learning Algorithm

Issued Dec. 30, 2025. Reinforcement-learning system for automatically adjusting bidding policy functions in content presentation campaigns. The practical version of this work was automated campaign control: translating advertiser objectives and realized outcomes into bidding policy updates that could operate in a real ads marketplace. Assignee: Maplebear Inc.; inventors include Jonathan Gu.

Issued patent Reinforcement learning Automated bidding

Broader patent portfolio

Additional filed or submitted work covers double-wide ad auctions, synthetic treatment effects for targeting, dynamic offer targeting, retailer classification using sales data and LLMs, critical-item fulfillment optimization, customer-centric replacement agents, and item display decisions in e-commerce.

Writing

Selected writing.

Writing is where I try to make the underlying logic visible: what shipped, what the tradeoffs were, what it cost, and what actually seems true.

Contact

Get in touch.

I'm usually interested in applied ML, agent infrastructure, marketplace design, local software, and projects where the system has to make real decisions.

Reach me directly

Use whichever channel is easiest. Email is best for thoughtful notes about projects, collaboration, or roles that actually fit.

Email LinkedIn Twitter / X GitHub Resume (PDF)