← Selected work
01 / 03Voice AI · Lead conversion

Riya voice platform

A voice agent designed to know when the conversation should no longer belong to the model.

LLMVoiceHuman in the loopDrift
System architecture / simplified
  1. 01Leadintent + context
  2. 02DialogueHinglish voice loop
  3. 03Confidence gatesignal + sentiment
  4. 04Next actionconfirm or nudge
Low confidence → human sales handoff

Stanza Living's inbound volume reached tens of thousands of leads a month. Response latency hurt conversion, while sales teams spent valuable time on repetitive qualification and visit-confirmation calls.

The obvious automation was not the whole problem.

A conventional English-first bot could automate calls but not the real interaction. Leads code-switch mid-sentence, intent becomes ambiguous, and forcing the model through a low-confidence exchange creates a worse handoff later.

The reliable product was not a more persistent bot. It was a system that could recognise uncertainty early and give the conversation to a person with enough context to continue well.

Build the operating loop, not only the intelligent step.

  1. 01

    Mapped the lead-confirmation journey around user intent rather than a fixed call script.

  2. 02

    Designed for Hindi-English language drift instead of treating code-switching as an edge case.

  3. 03

    Used confidence and frustration signals to route uncertain calls to human operators.

  4. 04

    Made the handoff an explicit product state, with context preserved for the sales team.

04 / Outcome

The platform now covers the top-of-funnel conversation at scale, allowing the team to spend less effort on call volume and more attention on high-intent conversations.

40Kleads handled each month
60 / 40Hinglish code-switching pattern
HITLconfidence-based routing

What stayed after shipping.

A voice agent earns trust by failing legibly. The human fallback is not evidence the AI is weak; it is evidence the product understands the cost of pretending certainty.