Bitontree Case StudiesAll work
← All case studies

Candor IVF Center / Radha Hospital

IVF Call Analytics - AI-Powered Call Analytics Platform for IVF Clinic Operations

Bitontree delivered a dedicated AI-powered call analytics platform for IVF clinic operations.

IVF Call Analytics - AI-Powered Call Analytics Platform for IVF Clinic Operations

About The Client

Candor IVF Center / Radha Hospital is a fertility healthcare provider that handles a high volume of patient calls across lead inquiries, follow-ups, visit coordination, camp campaigns, birthday calls, complaints, and operational support.

The client needed a system that could convert call recordings and MCube call activity into structured operational intelligence: call summaries, agent quality scores, missed script steps, high-intent lead signals, conversion flags, and daily performance reports.

Project Detail Information
Client Candor IVF Center / Radha Hospital
Industry Fertility Healthcare, IVF Clinic Operations, Patient Call Center
Location India
Duration 20 days
Project Scope IVF call analytics backend, dashboard frontend, transcription pipeline, AI call analysis, and agent reporting
Services Used AI Automation Development, Custom Software Development, Healthcare Call Analytics, Workflow Automation, Cloud Deployment
Engagement End-to-end product engineering across frontend, backend, call processing, and AI analysis
Live Link https://candor-ivf.netlify.app/login
Deployment Backend on AWS, frontend on Netlify

The Objective

Candor IVF Center / Radha Hospital wanted to reduce the manual work required to review patient call quality and agent performance.

The goal was not just to transcribe calls. The platform needed to understand real IVF call center conversations, identify the call type, score agent behavior, detect missed conversion opportunities, protect patient privacy, and give managers a reliable dashboard for coaching and performance review.

The platform needed to:

  • Receive MCube call webhooks and store call metadata automatically
  • Process answered call recordings through an asynchronous worker pipeline
  • Transcribe calls using Sarvam speech-to-text
  • Mask patient and agent personally identifiable information before AI analysis
  • Analyze calls against IVF-specific scripts and quality criteria
  • Detect lead calls, follow-up calls, visit calls, camp calls, birthday calls, complaints, wrong numbers, and no-content calls
  • Score greeting, empathy, accuracy, clarity, call-type handling, appointment handling, closing, and overall quality
  • Flag missed conversions, weak follow-up, wrong information, wrong tone, unresolved complaints, financial hesitation, and missed escalation
  • Surface high-intent patients from call content
  • Provide managers with call volume, call logs, per-call analysis, and agent-level reports
  • Generate daily agent reports with cached and incremental update support

The Challenge

IVF call center conversations are sensitive, multilingual, and operationally important. A weak greeting, missed camp explanation, unclear pricing response, poor appointment close, or delayed follow-up can reduce patient trust and conversion.

Before automation, managers had to rely on manual call sampling and subjective review. That made it difficult to know which agents were following scripts, which patients were high intent, and where conversion opportunities were being lost.

The key challenges included:

  • High call volume made manual recording review impractical
  • Calls included Gujarati, Hindi, English, and mixed-language conversations
  • IVF conversations contained sensitive patient and medical context
  • Raw transcripts could not be sent to AI providers without privacy protection
  • Generic sentiment analysis was not enough for IVF-specific quality review
  • Call scoring needed to account for lead calls, follow-ups, visit calls, camp calls, birthday calls, and complaints differently
  • Managers needed plain-English coaching insights, not internal AI labels
  • MCube webhooks needed reliable ingestion, duplicate handling, and queue triggering
  • Long-running recording processing needed to happen outside the webhook request path
  • Agent reports needed accurate operational totals instead of relying only on LLM-generated numbers
  • Daily reports needed caching and incremental updates to avoid unnecessary regeneration

Candor IVF needed a production-ready call intelligence platform that could handle real clinic conversations while keeping patient privacy, agent coaching, and operational accuracy at the center.

Our Approach

Bitontree designed IVF Call Analytics as a full-stack call intelligence platform: a FastAPI backend for call ingestion and reporting APIs, a Celery/Redis worker pipeline for recording processing, and a Next.js dashboard for authenticated manager review.

The system was built around PostgreSQL models for agents, calls, users, refresh tokens, and generated agent reports. MCube became the source of call metadata, Sarvam handled transcription, the masking layer protected sensitive transcript content, and configurable LLM providers handled structured IVF-specific analysis.

Our approach focused on:

  • Treating MCube webhooks as the operational source of call records
  • Separating fast webhook ingestion from slow audio and AI processing
  • Masking PII before sending transcripts to LLM providers
  • Using domain-specific prompts for IVF call quality instead of generic call summaries
  • Combining AI-generated qualitative insights with deterministic backend calculations
  • Building manager-facing dashboards around daily review, call logs, call analysis, and agent reports
  • Deploying the backend on AWS and the frontend on Netlify for a clean production split

Core Features Developed

MCube Call Ingestion

We built a protected FastAPI webhook endpoint that receives MCube call data, validates payloads, creates or updates agent records, stores call metadata, detects duplicate call IDs, and decides whether the call should enter the AI processing queue.

Asynchronous Recording Processing

Answered calls with recording URLs are processed through Celery workers backed by Redis. The worker downloads and preprocesses audio, creates Sarvam transcription jobs, uploads the recording, starts transcription, polls for completion, finalizes the transcript, and updates call status.

Sarvam Speech-To-Text Integration

The backend integrates with Sarvam batch speech-to-text APIs for call transcription. The pipeline supports retryable requests, job polling, upload URL handling, status tracking, and failure logging.

Privacy-First PII Masking

We built a transcript masking layer that replaces patient names, agent names, phone numbers, email addresses, Aadhaar-style IDs, PAN-style IDs, dates of birth, and relevant organization references before LLM analysis.

The masking combines known call values, regex patterns, IVF-domain dictionaries, and Hugging Face IndicNER support for Indian-language transcripts.

IVF-Specific Call Analysis

Each masked transcript is analyzed against IVF-specific scripts and call center expectations.

The analysis captures call type, language, agent scores, score reasons, missed script steps, flags, high-intent lead criteria, patient intent, patient sentiment, appointment booking, callback commitment, camp offer handling, query resolution, conversion relevance, and transcript quality notes.

Configurable LLM Provider Layer

The backend supports Claude or Gemini through an environment-based provider setting. It stores provider, model, token usage, estimated cost fields, generation status, and error information with each analysis or report.

Dashboard And Call Summary

We built an authenticated Next.js dashboard showing call volume, inbound/outbound split, answered calls, customer and executive disconnects, average duration, total talk time, day-wise trends, answered/missed trends, direction breakdown, and recent answered calls.

Call Logs And Per-Call Review

Managers can filter calls by date range, agent, direction, answer status, dial status, disconnected-by value, patient phone, and page through results. Each analyzed call can be opened in a detail drawer showing scores, call type, flags, and analysis output.

Agent Report Generation

The agent report module generates date-range performance reports for selected agents. Reports include call volume, analysis scope, call type breakdown, performance scores, conversion outcomes, flags, high-intent lead analysis, best call, worst call, strengths, areas to improve, and overall assessment.

Cached And Incremental Reports

Generated reports track source call IDs and call versions. If the source calls have not changed, the backend returns the cached report. If new calls were added, the system can update the report incrementally instead of regenerating everything from scratch.

Daily Agent Report Scheduler

We added an APScheduler-based daily report cron with PostgreSQL advisory locking so scheduled report generation can run safely without duplicate concurrent cron runs.

Authenticated Frontend Proxy Flow

The frontend uses Next.js API routes for login, logout, refresh, session reads, agent fetching, call summaries, call logs, call analysis, and agent reports. Tokens are kept in HTTP-only cookies and protected dashboard routes are handled through the frontend proxy flow.

Technology Stack

Layer Technology
Frontend Next.js 16, React 19, TypeScript, Tailwind CSS 4
Charts And UI Recharts, Lucide React
Backend API Python 3.12, FastAPI, Pydantic
Database PostgreSQL 16
ORM And Migrations SQLAlchemy, Alembic
Queue And Workers Celery, Redis
Scheduler APScheduler
Call Source MCube webhooks
Speech-To-Text Sarvam STT
AI / LLM Claude or Gemini, configurable by environment
PII Masking Hugging Face IndicNER, regex rules, known values, IVF-domain dictionaries
Authentication JWT access and refresh tokens, HTTP-only frontend cookies
Deployment Backend on AWS, frontend on Netlify
Local Runtime Docker Compose
Package Management npm for frontend, uv for backend

Architecture Overview

[MCube Call Platform]
        |
        v
[FastAPI MCube Webhook]
        |
        v
[PostgreSQL - Agents, Calls, Reports]
        |
        v
[Redis + Celery Worker Queue]
        |
        v
[Recording Download + Audio Preprocessing]
        |
        v
[Sarvam Speech-to-Text Job]
        |
        v
[Transcript Finalization]
        |
        v
[PII Masking Layer]
        |
        v
[LLM Call Analysis - Claude or Gemini]
        |
        v
[Stored Call Summary + Scores + Flags]
        |
        v
[Next.js Dashboard on Netlify]
   |             |              |
   v             v              v
[Call Summary] [Call Logs] [Agent Reports]
                                  |
                                  v
                     [Daily APScheduler Report Cron]

The Result

Bitontree delivered a dedicated AI-powered call analytics platform for IVF clinic operations.

The system brings MCube call ingestion, recording processing, transcription, privacy masking, LLM analysis, manager dashboards, call logs, and agent reporting into one workflow.

Outcome Result
Call ingestion MCube webhook activity stored directly in the analytics database
Recording workflow Answered recorded calls processed through Celery and Redis workers
Transcription Sarvam STT converts recordings into transcript text
Privacy protection Sensitive patient and agent information masked before LLM analysis
Call scoring IVF-specific scoring across greeting, empathy, accuracy, clarity, appointment handling, closing, and overall quality
Lead intelligence High-intent patients surfaced from call content
Conversion visibility Missed conversions, weak follow-up, financial hesitation, and appointment outcomes flagged
Manager dashboard Call volume, trends, logs, scores, and per-call review available in one interface
Agent reports Date-range performance reports generated for agent coaching and management review
Report reliability Cached and incremental report generation supported through source-call tracking
Production deployment Backend deployed on AWS and frontend deployed on Netlify

Business Impact

The platform gives Candor IVF Center / Radha Hospital a repeatable way to improve call center quality without manually reviewing every recording.

The final solution helped the clinic:

  • Reduce manual effort required to review call quality
  • Identify high-intent patients who may otherwise remain hidden inside recordings
  • Detect missed follow-up and missed conversion opportunities faster
  • Coach agents using real call scores, missed steps, flags, and examples
  • Improve consistency around greeting, empathy, clarity, appointment ownership, and closing
  • Protect sensitive patient information before AI analysis
  • Compare agent performance across calls and date ranges
  • Connect call volume, quality, patient intent, and conversion outcomes in one dashboard
  • Generate daily reports for leadership and manager review

For the business, the biggest shift is moving from reactive call sampling to proactive call quality control. Managers can review patterns early, coach agents with evidence, and act on patient conversion signals before opportunities are lost.

Development Timeline

Phase Deliverables
Phase 1 MCube webhook ingestion, agent and call database models, API key protection, and backend foundation
Phase 2 Celery/Redis processing pipeline, recording handling, Sarvam STT integration, and transcript finalization
Phase 3 PII masking, IVF-specific call analysis logic, configurable Claude/Gemini LLM layer, and per-call summary storage
Phase 4 Next.js dashboard, protected login/session flow, call summary metrics, call logs, filters, and call analysis drawer
Phase 5 Agent report generation, deterministic report metrics, cached and incremental reports, daily scheduler, AWS backend deployment, and Netlify frontend deployment

Why This Project Stands Out

Candor IVF needed more than a basic call recording dashboard or a generic AI call summarizer. IVF patient communication has its own operational realities: sensitive fertility conversations, multilingual calls, lead follow-ups, camp promotion, appointment ownership, financial hesitation, privacy protection, and high-intent patient signals.

Bitontree built IVF Call Analytics around that real clinic workflow. The result is a healthcare call intelligence platform where AI helps structure and analyze patient conversations, while managers keep clear visibility into agent performance, call quality, follow-up gaps, and conversion opportunities.

Publishing Notes

  • TODO: Confirm preferred LLM provider and operating cost model before final go-live.