About

Now

Learning

Interaction models

Researching

Mech interp (GEAR program)

Project

MNIST (Rust)

Based in

India

I build and research interaction models: AI that can listen, think, and speak at the same time.

Before that, I was in systems programming, building interpreters, language parsers, and VMs in Rust. Along the way, I have built a few language parsers and interpreters. But the most important lesson I learned is how to take a complex problem, break it down into tiny chunks, solve them independently, and put them back together to solve the whole problem.

Recently, I got into machine learning. I built an MNIST classifier model from scratch in Rust with no ML libraries. This was a great learning experience that taught me about the forward pass, backpropagation, gradient descent, and the whole training loop.

Along the way, I discovered the interaction model from Thinking Machine Lab: A model that listens, thinks, and knows when to stop and when to respond. I am really impressed by its demo. This sparked my curiosity to explore the whole world of full-duplex and interaction models.

Now, I'm learning, building, and researching these interaction models. Starting with the data pipeline, model architecture, SFT, RL, inference, alignment, and safety, I explore the complete vertical stack of these models.

This October, I joined GEAR (Gradual Empowerment Alignment Research Program) to learn ML research and mechanistic interpretability. I share my learnings, findings, mistakes, and thoughts here, on Twitter, on LinkedIn, and sometimes on YouTube.