Hinglish voice agent for lead qualification

In-house build. The demo is being upgraded, so it appears here as a case study only.

Label: IN-HOUSE BUILD
Result: Demo being upgraded; case study only for now
Stack: FastAPI · Redis · Groq · Saaras · Bulbul

Problem

Qualify insurance leads by phone, in natural Hinglish, within IRDAI-aware conversation rules.

Approach

A speech-to-text and text-to-speech pipeline on Groq (Llama 3), Saaras and Bulbul.
Redis session state, so each call keeps its context turn by turn.
Two models: one writes the reply, the other extracts the lead details.
Prompts tuned for a faster first response.

Architecture

- Caller Phone call in Hinglish
- Saaras Speech to text
- Reply model Groq · Llama 3
- Lead extraction Groq · structured fields
- Bulbul Text to speech
- Redis Session state per call

Speech in, speech out, with two models working side by side on each turn. Lead extraction, the part that produces the business result, is drawn in red.

Result

The demo is being upgraded, so it appears as a case study only for now.