Saloni Gandhi

Saloni Gandhi is a senior AI engineer at LinkedIn. She helped develop semantic search for jobs and people and, with Ziwei Guo and colleagues, studied how well LinkedIn’s career agent understood members’ job preferences. One mistake was easy to miss: the agent could suggest a title that fit someone’s résumé while overlooking the work that person had actually been searching for. An automated judge could give the agent credit for that answer.

From recommendation systems to search

At Cornell, Saloni earned two master’s degrees, specialized in recommendation systems, and served as a teaching assistant in computer vision and deep learning. She also worked on a model that combined vision and language at the Robotics and AI Institute. Cornell offered her a place in its doctoral program, but she found she preferred building products people use.

At Twitter, Saloni led a change to how the platform ranked replies in conversations. Her team used spam signals to move spammy replies farther down a thread, reducing daily views of those replies by about a million, as her colleague Tyler Romero recalled. At LinkedIn, she co-authored the 2026 paper Semantic Search At LinkedIn. The team used embeddings to find jobs and people whose descriptions matched the meaning of a query even when the wording differed. To rank those results within the time available for a search, the team trained a compact model to reproduce the judgments of larger models.

When the judge gets a preference wrong

Saloni and her colleagues compared the career agent’s automated judge with roughly 180–200 human reviews. Besides accepting plausible but incomplete job preferences, the judge sometimes gave its lowest score to a partly correct answer. The team clarified when to award partial credit and which variations in a job title or work arrangement preserved its meaning. A difference in the capitalization of “full-time,” for example, had been enough to cost the agent points. After the revisions, agreement with human reviewers rose from 57% to 64%.

Saloni and her colleagues found a separate problem with the model that diagnosed low scores. Its list of possible causes left out errors by the judge itself. The model could repeat the judge’s reasoning and blame the scoring rules even when the judge had misapplied them. That distinction mattered because the diagnosis guided changes to the agent’s instructions. The team tested proposed edits on held-out examples and kept human review in place when it shipped its first prompt change proposed by GEPA.