Instacart Ads Automated Bidding
Budget-based marketplace ads bidding, return estimation, pacing, simulation, and production rollout.
PhD Economics, UCLA · Member of Technical Staff, Applied Science at Coframe
Jonathan Gu
Economics, machine learning, and decision systems.
I build production systems where a model has to make a real decision, leave enough evidence to audit it, and improve without outrunning its guardrails.
Work
Work worth inspecting.
The production marketplace systems come first. Current independent work carries the same discipline into agent memory, adaptive experimentation, neighborhood food, and local coordination.
Instacart Growth Machine Learning System
Causal targeting from experiment outcomes: expected incremental effect, budget allocation, and policy evaluation.
Instacart Inventory Intelligence
Item-store availability and fulfillment ML with noisy catalog, store, shopper, and operational signals.
ocbrain
Local, source-backed shared memory for Codex, Claude Code, OpenClaw, and other MCP clients. It retrieves bounded context from private history, expands exact sources on demand, and records whether the context actually helped.
Building now · Live
Bountiful Garden
A phone-first neighborhood marketplace for homegrown food. Neighbors can make a garden, post what is ripe, discover nearby harvests, message each other, and arrange a share or trade.
The Decision Loop / Coframe Brain
A working book about the system between generation and allocation: propose safely, assign traffic, log what happened, join outcomes, and decide what is worth learning next.
RackeTapp
A trilingual, mobile-first pilot for finding players to join an already-booked Hong Kong tennis court, splitting the fee, and getting lawful availability alerts—without automated booking.
Stringstep
A collaboration system for turning ambiguous work into inspectable plans, evidence, and clean handoffs without hiding the reasoning that produced them.
Policy Agent
A retrieval-first policy assistant that keeps source documents, jurisdiction, and privileged operations visible instead of answering from a generic policy-shaped memory.
Pelican
An inspectable forecasting and decision dashboard built around calibration, uncertainty, and the difference between a model score and a decision worth taking.
Links
Writing, papers, and proof.
Straight to the useful stuff.
ResumePDF
Ads Automated BiddingTechnical writeup
Growth MLCausal incentive allocation
Issued patentAutomated policy-function adjustment
School choiceResearch PDF
Returns to schoolingResearch PDF
Reinforcement learningResearch PDF
Temporal Fusion TransformerExplainer
Bountiful GardenLive product
ocbrainInstall and verify
The Decision LoopCoframe Brain manuscript
RackeTappTrilingual tennis-court pilot