I'm Roy Benjamin.

I lead conversational analytics for Tableau at Salesforce: the layer of Tableau Next that lets people ask questions of their data in plain language and get answers they can defend in a meeting. It is live with dozens of enterprise customers, which means my work is graded every day by people who never saw the demo. 1

Making a language model reliably useful in front of enterprise data is mostly unglamorous work: evaluation, trust, and semantics. I have learned that the unglamorous parts are where the product actually lives. Enterprise data is simply this chapter's version of a problem where being wrong is expensive.

I got here the long way around. It started in undergrad with two projects that never quite let go of me: a DIY drone we built and programmed for 3D reconstruction, and GANs that pieced shattered mosaics and frescoes back together from their surviving fragments. From there to physical design on Apple's silicon team, where a chip ships exactly once. 2 A Master's in computer science at the Technion followed: a first-author paper at CIKM 2022 on graph neural networks for molecular property prediction, advised by Kira Radinsky. And at Gloat, applied AI research, transformer ranking models and responsible AI, I made the crossing from research into product.

I still keep a foot in the research world, as a NeurIPS ethics reviewer and a WSDM reviewer, and I am completing an MBA at Imperial College London: the business half of the same equation.

The problems I keep coming back to are the ones where software has to act in the physical world: state estimation, computer vision, systems that hold a belief and update it as evidence arrives. The current one is nano-kalman, a Kalman filter written from scratch, tracking a noisy ball live in the browser. 3 Before that it was drones and broken frescoes; there will be a next one.

Off hours I host The Fine Print (האותיות הקטנות), a Hebrew podcast where I sit down with people doing extraordinary things (researchers, founders, industry professionals) and ask about the details that usually go untold. 4 16 episodes and more than 3,000 downloads since July 2024.

I have been explaining complicated things to sharp audiences for a long time: about 600 students over two years as a CS teaching assistant at the Technion, with multiple Excellence in Teaching awards, 5 lectures for SpaceIL since 2015, and mentoring at Startup Nation Central.

This page is the short version. For the rest: email hello @ this website, or find me on LinkedIn, where I am most active, and GitHub. Organizing an event? Include the date and the audience, and I will get back to you quickly.