My current focus is applied AI — training small language models and distributed inference, getting capable models to fine-tune and serve on the hardware you actually have.
I'm co-founder of Prashnam, India's first AI-powered telephonic opinion platform — its survey data has been cited by The Economist and The New York Times — and I lead product and technology for Atrium Solutions, building a forecasting and decision-intelligence platform. Over two decades I've held senior product roles across Times Internet, India Today, IndiaMART, and Samsung.
Lately my work has centered on small language model training and distributed inference — getting capable models to fine-tune and serve on the hardware you actually have.
Most of the rest is a collection of single-file web apps that run entirely in your browser — no accounts, no servers, no data leaving your device.
The frontier work pulls two ways at once: pushing model inference all the way into the browser tab (text-diffusion language models in kohra), and, when a model is simply too big for a tab, serving it on cheap spot GPUs that survive being reclaimed mid-run (Cairn, with its recovery core rebuilt from scratch in Rust as warmswap). A newer thread is tooling for coding agents — vise freezes what a codebase does and gates an agent's changes against it. Everything public is open source.
Frontier-model inference on cheap spot GPUs — the kind the cloud can reclaim from you at any second. A 671B open model striped by layers across single-GPU spot boxes; when a box dies mid-generation its slice moves to a warm spare and the token stream resumes bit-identical in 0.93 s. Proven live on a real fleet.
On-device voice commerce for online stores. Shoppers talk, search, ask about a product, and hear the answer read back — all in their own browser, on a per-merchant brand-tuned model. Zero inference cost per shopper.
In development · private