Yexoria logoYexoria logo
Back to Projects

Case study

2026AI

AI voice intelligence platform

Businesses hold thousands of customer conversations, but most of what's said stays trapped in audio. We helped an insurer build two connected AI capabilities: one that listens to every recorded call, and one that picks up the phone itself.

Client
Insurance company
Year
2026
Services
Mobile engineering · UX research
Platform
Voice agent, analytics, dashboard
Call waveform coloured by customer emotion with a live transcript, next to an AI voice agent making a feedback call

The problem

Every day an insurer has thousands of conversations with its customers. Each one ends, gets saved as audio, and is never heard again.

One call, start to finish

The journey of a single conversation

One call, start to finish

The journey of a single conversation
01 / 05 · THE CALL

It starts as sound.

A recorded call-centre conversation. Valuable, but locked in audio and far too many to review by hand.

02 / 05 · THE WORDS

Speech becomes text.

Real-time speech-to-text, tuned for Indian customer interactions, turns the call into something the system can read.

I've called twice about my claim and nobody got back to me. I just need to know when it will be settled. The agent was helpful today, thank you.

sample transcript
03 / 05 · THE FEELING

Text reveals emotion.

Sentiment, emotion and call quality, scored on the conversation instead of a sampled checklist.

I've called twice about my claim and nobody got back to me. I just need to know when it will be settled. The agent was helpful today, thank you.

negativepositive
04 / 05 · THE INSIGHT

Emotion becomes a decision.

Agent performance and dissatisfaction patterns land on a dashboard the business can act on.

  • Repeated claim-status callspattern
  • Frustration rises after holdsentiment
  • Agents needing coachingquality
  • Customers to call backaction
05 / 05 · THE CALL BACK

Then the system calls the customer.

An AI voice agent built on LiveKit reaches out on its own, collects structured feedback, and sends it back to the start.

“Hi, I’m calling about your recent claim. Do you have two minutes?”

sound
loop

●Under the hood

Two lanes, one loop

LANE A · ANALYSELANE B · ENGAGE ←
A1Recorded calls
A2Speech-to-text
A3Sentiment & emotion
A4Pattern detection
B1Insight dashboard
B2AI voice agent
B3Proactive calls
B4Structured feedback
node 1 / 8

Recorded calls

Call-centre conversations that were previously reviewed only by sampling.

solves → Valuable conversations trapped in audio.

↺ Feedback returns to lane A, so the loop keeps learning

What we did

From a concept to a working loop

We helped conceptualise the ecosystem and then built it, with research, design and engineering in one team.

UX research

Understood how call-centre teams and managers use feedback today.

Product design

A dashboard that turns AI output into business-friendly insight.

Voice AI

Speech-to-text, text-to-speech and a real-time voice agent.

Mobile engineering

Production apps and audio pipelines in Python.

  1. Discover

    Listened to real calls with the client team and mapped where feedback was lost.

    • call audit
    • stakeholder interviews

  2. Conceptualise

    Designed the two-lane system: historical analysis plus proactive engagement.

    • system concept
    • voice persona

  3. Build

    Built the conversation intelligence pipeline and the LiveKit voice agent.

    • audio processing
    • sentiment models
    • LiveKit

  4. Tune

    Tuned Sarvam AI speech for Indian customer interactions until conversations felt natural.

    • Indian languages
    • conversation design

  5. Visualise

    Shaped the dashboard experience that turns AI output into decisions.

    • dashboard concept
    • insight design

Built with

The stack behind it

VOICE
LiveKitSarvam AISpeech-to-textText-to-speech
INTELLIGENCE
Conversational AISentiment & emotion analysisPattern detection
ENGINEERING
PythonAudio processingMobile apps

Built for India

Speech models tuned for Indian customer interactions.

Human in the loop

Managers see the reasoning and decide what to act on.

One continuous loop

Listening and engaging feed the same intelligence layer.

Scroll to switch the system on

BEFORE

Conversations went unheard.

  • ✕Valuable conversations were trapped in audio, too many to review by hand.
  • ✕Customer emotion couldn't be measured with traditional call metrics.
  • ✕Feedback was reactive and depended on surveys and complaints.
AFTER YEXORIA

Every call feeds the loop.

  • ✓Far less dependence on manually reviewing large volumes of calls.
  • ✓Sentiment and emotional signals surfaced from conversations.
  • ✓A proactive, AI-driven feedback loop running alongside historical analysis.

The product

Customer emotion, in business language

Customer insight dashboard with sentiment trend, top dissatisfaction drivers and the feedback loop
Sentiment trends, the biggest drivers of dissatisfaction and the feedback loop on one dashboard.

Impact

What the platform does

2connected AI systems: call analysis and a voice agentsystem design
Real-timespeech-to-text and text-to-speechsystem capability
Indiancustomer interactions, via Sarvam AI speech modelssystem capability
1 loopanalyse → understand → engage → learn → improvesystem design
OUTCOME 1

Teams no longer depend on manually reviewing large volumes of calls.

OUTCOME 2

Emotion that was hidden in audio is visible on a dashboard.

OUTCOME 3

Feedback is collected proactively, in one continuous loop.

Could this work for you?

The same loop fits any business that talks to customers

Banking & insurance

Claims, collections and service calls.

Healthcare

Patient follow-ups and appointment feedback.

E-commerce

Support calls, returns and delivery issues.

Real estate & auto

Lead calls, site visits and after-sales.

Next case studyAI call conversion engineAn AI system that listens to every sales call, scores it for intent the moment it ends, and routes the hottest leads straight back to reps, synced to the CRM and queryable through Claude and ChatGPT.Read next