# About VARTA

## What we believe

A voice agent that only has to survive a short demo in a quiet room is a different, easier
problem than one that has to survive a long call in Hinglish with a caller on a scooter who
interrupts, changes their mind, and gets frustrated. We built VARTA for the second problem, on
the belief that the language model should be the exception in a call flow, not the engine running
it: deterministic logic decides first, and the model is called only when nothing else can,
constrained to a shortlist rather than left to compose freely.

That same belief shapes what we chose to build rather than buy or bolt on: caching that keeps
working once a call is personalised instead of collapsing the moment it is, a per-call cost
ledger instead of a blended rate, an engine with no language-specific text so there's no English
default to slip back into, a regression gate so a change in behaviour is measured rather than
anecdotal, and empathy handling written down as explicit policy rather than left to a prompt's
judgment.

## Built to be run, not just demoed

**humanise.ai** runs as a live, multi-tenant SaaS product on this exact engine today, handling
real calls in production -- this is not an architecture we believe in on paper. If your
organisation needs to run that same engine inside infrastructure you control, that's what our
[enterprise licence](/enterprise) is for.