About
A short note about me.
I am Kushan, an AI engineer and builder from IIT Madras. I like systems that sit close to real use: calls, browsers, language, images, geometry, and the small frictions that decide whether a tool actually helps.
I am most interested in capability. Not as a slogan, but as a test: after this system exists, what can someone understand or do that was harder before?
The work
Systems before slogans.
At Sarvam, I worked on production voice-agent and LLM systems: inference, orchestration, retrieval, interruption handling, telephony, and the many practical details that make an agent usable in real conversations.
Around that, I have spent time on browser agents, multimodal translation, medical-imaging diffusion, point-cloud completion, and 3D reconstruction. The recurring question is practical: what representation lets a system act more honestly and usefully?
Background
Biology, complex systems, and taste.
I studied Biological Engineering and Complex Systems and Dynamics at IIT Madras. That mix still shows up in my taste: I like systems with feedback, noisy data, embodied constraints, and a gap between what is measured and what matters.
I also spent a lot of college inside music and theatre: leading the music club, playing guitar, forming bands, performing, and learning how much good work depends on rhythm, trust, taste, and rehearsal.
Now-ish
What I am circling.
- Browser agents that see better state.
- Voice agents that stay useful in real conversations.
- Cleaner data loops before bigger model loops.
- World models, robotics, geometry, and physical representations.
- Tools that make people more capable without making them feel smaller.
Elsewhere
A few public traces.
For now, the public links here are intentionally sparse: a talk, a paper, and product context for work I have been close to. More links can come once they are worth keeping around.