About

I’m Vaibhav Nakrani, a Machine Learning and Agentic AI engineer. What I do, in one line: read modern research, pull out the idea that matters, and implement it against the hardest version of the problem in front of me.

That last part is the whole game. A lot of ML lives in the gap between a paper that works on a benchmark and a system that holds up at 4K @ 30 FPS in front of real users. I’m comfortable on both sides of that gap, and I think the people who can move between them, keeping the science honest while making the engineering ship, are the ones who actually solve the tough problems.

A few of those problems I’ve taken on:

The throughline is a balance between theory and engineering: I’ll dig into why a model behaves the way it does (interpretability with Captum, attention maps, occlusion studies, rigorous data cleaning) and then own the unglamorous path to production on AWS/GCP. My core stack is PyTorch, the YOLO/SAM/DETR family for vision, and LangGraph with the OpenAI/Anthropic SDKs for agents.

A recent review put it better than I can:

Vaibhav showed outstanding results on image recognition tasks. He has a remarkable balance between science, theory, and engineering. He is proactive and suggests well thought-out solutions. Communication is excellent. Recommend working with him strongly.

Why this blog

Most ML explanations either hand-wave the intuition or drown you in notation. I write the version I wish I’d had, building ideas up one step at a time until they click, with the math kept honest but never used as a smokescreen. Expect posts where I take a piece of modern research and rebuild it from the ground up: diffusion, agents, computer vision, and whatever I’m currently figuring out.

Elsewhere