About Me
Independent research in mechanistic interpretability,
efficient inference and intelligent systems.
Hello there.
I'm S. Khavin. Thanks for visiting my little corner of the internet.
My journey started when I was seven, building command-block and redstone machines in Minecraft. I've always liked taking things apart, figuring out how they work, and then trying to build something better.
Today, my research sits at the intersection of mechanistic interpretability, efficient inference, and computer vision.
I love building things. My first research project came from a problem I ran into while building AgentFox, an agentic browser. I wanted to ship it as a self-contained system: Python environment, packages, and a VLM running locally. Then I ran into the problem of video understanding: processing long videos with transformer-based vision models can become extremely expensive as the amount of visual context grows.
So I built Trinetra, a temporal architecture designed to reduce the amount of visual information that needs to be processed while preserving the information necessary for later queries.
But that led to another problem.
If we can trace how a model computes, head by head, layer by layer - can we use that map to make inference dramatically cheaper?
That question eventually led me toward mechanistic interpretability.
My current research explores a simple idea: if we can trace how a model computes, can that map tell us what computation actually needs to be preserved? And if we understand those structures well enough, can we use that understanding to make inference cheaper, faster, and more controllable, without changing the model's weights?
I also don't completely buy the idea that AI is not a black box. We can inspect the weights. We can inspect the activations. We can intervene on individual components. But that doesn't necessarily mean we understand why a model produces a particular computation.
Beyond efficiency, I'm interested in comparing these artificial systems with biological intelligence, not because I expect transformers to simply resemble the human brain, but because the comparison can help us ask a more fundamental question:
How intelligent are today's AI systems, really?
RESEARCH QUESTIONS
The questions I keep returning to.
RESEARCH RECORD
PROFESSIONAL RECORD
EDUCATION
Sathyabama Institute
RESEARCH METHODS
SELECTED DISTINCTIONS
OUTSIDE THE LAB
I still like building things that have nothing to do with research.
Minecraft was probably the beginning. Then came web projects, little tools, hackathons, and an unreasonable amount of time spent thinking about how interfaces should feel.
Somewhere between all of that, I started taking research seriously.