Ziwei Guo

Ziwei Guo worked with Saloni Gandhi and colleagues at LinkedIn on a career agent that infers which jobs a member wants. As a staff software engineer in machine learning, Ziwei helped build a way to diagnose recurring errors in those inferences and test changes to the agent’s instructions.

Ziwei earned a bachelor’s degree at Fudan University in 2014 and a PhD in chemistry at Emory University in 2019. At Emory, Ziwei and adviser James T. Kindt used Monte Carlo simulations to study grain boundaries in two-dimensional crystals. In experiments, a patch of particles rotated just slightly out of alignment with the surrounding crystal disappeared fastest. In simulations, those patches lasted longest. By adjusting the simulated pressure, shrinking the patch and reproducing more closely how experimenters formed it, Ziwei and Kindt recovered the experimental trend. After Emory, Ziwei worked as a data scientist at Tubular Labs.

At LinkedIn, Ziwei used saved job searches to explain how the team’s user understanding agent handled evidence. If a title appeared repeatedly in a member’s saved searches, the agent could identify it as a preferred role, cite the searches and report its confidence. A separate model graded the inference. A low score could mean the agent had too little information, drew the wrong conclusion from the evidence or received an incorrect grade. That distinction mattered because changing the agent’s instructions to satisfy a mistaken grader could introduce an error.

The team grouped recurring failures by cause and used GEPA to propose prompt edits. They tested promising edits on held-out cases that had played no part in selecting them. The first such change reached production after quantitative evaluation and human review.