Vaibhav Hariram

CS and Geography at Berkeley. Most of my work sits on the correctness side of AI systems: evals, validation, and the tooling that tells you when a model is wrong. That turns out to be harder than making it smart.

I'm on leave from school this fall, working at Intel on AI/ML engineering for validation. This past summer I was there on power management controller and firmware validation, mostly pre-silicon.

Within geography, I tend to specialize in spatial machine learning and computational geoinformatics: routing algorithms, GIS pipelines, and how models represent physical space in the first place. Spatial intelligence is the direction I think matters most in AI right now. World models and physical AI are where I want my work to end up.

I also care about consumer software, and about applying AI to social problems people actually run into. Immigration is the one I've spent the most time on, with a few others I'm working toward.

Before this: research with Berkeley RISE Lab's Gorilla project on function-calling evals, so if a benchmark has ever told you your model's tool use is bad, some of that was my work. Paprika, a runtime policy and replay system for AI agents, running with 25+ teams now. Infra automating B2B content at Flywheel. I started on location management software at Railinc, which is where all of the spatial work comes from.

Recently: I've been getting into k-pop, looking forward to watching Chelsea disappoint me again this season, and getting back into seasonal anime. Planning a trip to Central Asia and convincing my roommates to get a cat for the apartment are also in the works.

Talk to me about any of it, anytime.