CoframeLive optimization
Jonathan Gu
Projects
ocbrainOpen source · v1.1.0
BountifulFresh nearby
RackeTappHong Kong · Live
PongPlayable now
Proof, not posture
Systems that survived contact with reality.
Good models are not enough. The work is defining the decision, measuring the counterfactual, building the control loop, and leaving a way to inspect what happened.
01
Personalization
Support-aware policy learning
A public walk-forward policy that combined empirical-Bayes shrinkage with gradient boosting and captured 69.6% of available personalization headroom.
02
Marketplace ads
Automated bidding at Instacart
Return estimation, budget pacing, auction-aware bid generation, simulation, rollout, monitoring, and an issued reinforcement-learning patent.
03
Causal allocation
Growth and inventory intelligence
Experiment-grounded uplift targeting and item-store availability systems built from noisy marketplace, catalog, shopper, and operational signals.
The lab
Small worlds with real rules.
Products, prototypes, and decision surfaces built to make an idea testable—not merely presentable.
Privacy-first product
StringStep
A visible safety Step that becomes a private, searchable timeline: start, record, retrace.
Project page ↗ Inspectable demoPolicy Agent
Retrieval-first policy answers that keep sources, jurisdiction, and privileged actions visible.
Open demo ↗ Decision dashboardPelican
An operator console built around calibration, uncertainty, and the difference between a score and a decision.
Inspect dashboard ↗Writing and research
Follow the reasoning.
Technical notes, papers, and primary evidence—the material behind the short version.