Rushil Jaiswal

I work on the layer that decides whether an AI system is right. In reinforcement learning a verifier scores the model’s work, so the verifier is what “good” means. Most people check whether it can be gamed. I check the other direction, whether it rejects work that is correct but different: vaudit found that three of five published grading strategies rank a blank screen above a correct implementation. The same question shows up in agent security. Airlock pays real invoices and was fooled once in 39 injection attacks, but never moved money, because payment destinations come from records instead of from anything a model read.

I’m a software engineering intern at a 140-person IT staffing firm and the sole engineer on their AI candidate-sourcing platform. It runs in daily production across about 80 open requisitions a month. Most matching tools rank résumés by keyword overlap with the job description. Mine ranks candidates on who actually got placed, and on what recruiters said about them.

Before college I started Wrapit Recovery, a patented athletic-recovery sleeve. 600 units and about $15K across five gym partnerships, with the manufacturing and support done myself.

Every project here lists what’s wrong with it. Most pages end with the part I couldn’t solve, and seven of the projects are dead.