VARTA
Log in

A voice-agent platform with a visible cost ledger

VARTA runs every call through seven explicit layers -- L1 through L7 -- instead of one opaque prompt. Repeated phrasing plays from cache instead of being regenerated, and only the turns that actually need reasoning reach an LLM. humanise.ai runs as a live, multi-tenant product on this same core today.

Try the browser demoSee your own cost
  1. L1cached — no model call100ms

    Caller: Hi, I need to reschedule my appointment.

    VARTA: Sure, I can help with that. Could I get your booking ID?

  2. L2cached — no model call150ms

    Caller: It's BK4521.

    VARTA: Got it, BK4521. What date works for you?

  3. L4LLM call650ms

    Caller: Could you check if Friday works, and if the technician can come after 5pm?

    VARTA: Friday after 5pm is available -- I've held that slot for you.

  4. L6cached — no model call100ms

    Caller: Thank you, that's a relief.

    VARTA: I understand -- glad we could sort that out for you.

  5. L7cached — no model call100ms

    Caller: That's everything, thanks, bye.

    VARTA: You're all set for Friday after 5pm. Have a great day!

Turns
5
Total cost
₹1.80(illustrative)
Cache hit rate
80%
A sample call, not a live benchmark -- four of five turns are cached (₹0 marginal cost); only the reasoning turn (L4) reaches an LLM.