I build AI that runs
where the compute isn't.

Modern models are built for people with spare compute, spare budget and spare bandwidth. A very large number of people have none of those, and they are not a rounding error — they are most of the world.

So my research is about compression. A crop-disease model at 1.6 MB that diagnoses offline on a farmer's phone. A classifier distilled down to a fraction of its teacher's compute without giving up its accuracy. And my company is the same instinct pointed at a different layer: giving a shop owner in Lahore the ad operations of someone paying an agency Rs 60,000 a month.

Same problem, one level up. Intelligence should work where the resources aren't.

You tell it what you sell and what you want. It writes the creative, builds the targeting, launches across Meta, Google, TikTok and X, then adjusts bids and budgets around the clock without anyone asking it to.

The Western tools don't fit this market. They assume card payments, when most orders here are cash on delivery and half the reported ROAS evaporates at the door. They assume the conversion lands on your website, when it actually happens in a WhatsApp thread the pixel never sees. They assume ad copy is written in one language. We're building for the market where all of that is the normal case, not the edge case.

It started as a final-year project at FAST called VoxCommerce. It's now incubated at NIC Islamabad, Cohort 5, run as a concierge service for local businesses while we watch every decision the AI makes — with self-serve access opening up shortly.