Agent systems and memory
OpenClawBrain is local-first AI assistant memory: retrieval, scoped memory, learned routing, calibrated abstention, replay, proof surfaces, and rollback.
Jonathan Gu
I build AI systems that turn model outputs into reliable decisions.
I am a machine learning engineer and PhD economist focused on applied AI systems: ads bidding, ranking, agents, memory, retrieval, forecasting, experimentation, and marketplace decisions that have to work under production constraints.
At Instacart, that meant building systems for bids, rankings, availability, causal measurement, and fulfillment. One major proof point is the first optimized ads automated bidding system, now covering roughly 70% of Sponsored Products spend, about $700M annualized.
Outside work, I build OpenClawBrain, Bountiful Garden, and Project Pelican: products about trust, memory, learned routing, and decisions under uncertainty.
Ads, agents, evals, calibration, rollout, and proof.
Ads & Ranking
Bidding, ranking, auction design, pacing, causal measurement, and marketplace decision loops.
Decision Science
PhD Economics, UCLA. BA Statistics & Economics, UC Berkeley. Causal thinking applied in production.
Builder Taste
Products that must decide: what to bid, what to remember, what to show, and when to route.
Research Systems
A balanced applied AI profile: ads, agents, and decisions.
My work sits at the point where models stop being demos and start becoming product behavior. The common thread is building systems that choose an action, measure the outcome, and stay understandable enough to debug.
Ads, ranking, and marketplaces
At Instacart, I built systems for bidding, ranking, forecasting, availability, fulfillment, and marketplace decisions where model outputs become real actions.
Evaluation and measurement
My economics background shows up in causal measurement, counterfactual thinking, calibration, experiments, monitoring, and rollout under uncertainty.
At Instacart
Production ML and marketplace decisions at Instacart.
My Instacart work spans automated bidding, ranking, causal marketing measurement, availability prediction, fulfillment, and production ML. A score is not enough. The system has to decide, run reliably, and measure what happened.
What I actually do
I own the full loop: problem framing, data, features, model and control design, serving, experiments, rollout, monitoring, debugging, and migration.
- Built production decision systems across ads, availability, fulfillment, and marketplace surfaces where predictions had to become reliable actions.
- Built Instacart's first optimized ads automated bidding system from scratch. It replaced manual product-keyword bids with weekly-budget bidding across roughly 70% of Sponsored Products spend.
- Started from ROAS equalization: estimate product and keyword returns, choose advertiser target ROAS, pace budgets, simulate bids, write bids, observe outcomes.
- Built causal marketing measurement and in-app decision systems where observed conversions were not clean labels.
- Built availability and fulfillment systems across noisy inventory signals, real-time LightGBM serving, calibration, rollout, and monitoring.
Automated bidding
Weekly-budget optimization, ROAS estimation, pacing, bid simulation, and bidding loops for a real ads marketplace.
Causal and marketplace measurement
Coupon and in-app marketing systems where incrementality and selection bias matter.
Availability and fulfillment ML
Models and serving paths for product-store availability, forecasting, calibration, and operational signals.
Selected Projects
Projects that show taste.
These projects share a pattern: turn fuzzy intent into systems that decide, preserve trust, and show their work.
Live product
Neighborhood software
Bountiful Garden
Bountiful Garden helps neighbors share backyard produce before it goes to waste. The goal is simple: make a local exchange happen.
It is founder-shaped work: product design, trust, and marketplace mechanics for ordinary neighborhood life.
Marketplace
Community
Full-stack product
Live product
Agent memory
v0.2.21
OpenClawBrain
OpenClawBrain gives OpenClaw agents local, inspectable memory. The current release turns the learning loop into a production route function, so the agent learns when memory should matter, not just what to store.
The spine is deliberately accountable: LLMs propose semantic meaning, code enforces trust boundaries, SQLite stores the graph and evidence, shadow replay tests candidate policies, route-policy-v3 serves only when confidence is calibrated, and v2/heuristics remain rollback paths. Current public release: 0.2.21.
Private R&D
Autonomous trading
Project Pelican
Project Pelican is a private autonomous options research system. It joins data, forecasting, risk controls, execution, and monitoring.
It is still internal. It reflects the same interest: systems that keep making measured decisions under uncertainty.
Quant finance
ML pipelines
Autonomous systems
Patents
Applied ML systems, backed by patents.
The patent portfolio reflects applied ML systems that made production decisions across bidding, marketplace design, causal targeting, retailer intelligence, and fulfillment optimization.
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. The practical work translated advertiser objectives and outcomes into bidding policy updates for 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, causal targeting, dynamic offers, retailer classification with LLMs, fulfillment optimization, replacement agents, and item display decisions.
Writing
Selected writing.
Writing is where I make the logic visible: what shipped, what failed, and what still seems true after the numbers come in.
Topics in Education and Labor (UCLA PhD Dissertation, 2020)
Three applied economics chapters on school grants, returns to education, and minimum wage policy.
Credible Inference for the Returns to Schooling (May 2020)
Bayesian and instrumental-variables analysis of heterogeneous returns to education.
Reinforcement Learning (Econometrics Field, July 2016)
An older paper behind part of my learning direction. Local PDF mirror.
Contact
Get in touch.
I am interested in applied ML, agent infrastructure, marketplace design, local software, and systems that make real decisions.
Reach me directly
Email is best for thoughtful notes about projects, collaboration, or roles that fit.