Oshri Naparstek

I trained my first neural network in high school, in Visual Basic, after reading Kurzweil. After three years in combat engineering, I spent 15 years doing what seemed like different things — applied math, signal processing, defense systems, distributed optimization. It took me a while to realize they were all the same thing.

During my postdoc at Washington University in St. Louis, I got curious about a reference in a widely-cited 2010 paper. It pointed to a 1967 paper nobody had read. I ordered a physical copy through the library. It turned out that 20 years of modern research had been unknowingly rediscovering what was already there. That experience shaped how I work — I always go to the original source, because fields forget.

At Rafael, I worked on reinforcement learning for defense — systems that had to learn in real time, with no room for error. At IBM, I co-authored Granite Vision, created the Real-mm-RAG benchmark, and manage the AI for Knowledge group. Currently a Principal RSM and Master Inventor.

Outside of research, I play bass, guitar, and piano — during my Master's I played bass in a wedding band to pay the bills. I'm a macro insect photographer and freediver. There's something about the small that interests me as much as the large — the complexity of a single compound eye, the patterns in coral. I'm also drawn to philosophy: questions about consciousness, understanding, and what it means to observe. I'm interested in a lot of things — probably too many — and I tend to spread across fields rather than dig into one. But I've found that the interesting ideas usually live at the intersections.