Case Study · AI Voice Agent · Outbound Automation

Autonomous AI Calling Agent Platform

Name: AI Calling Agent Platform
Author: Prabhdeep Singh

A production-grade outbound calling system that navigates IVRs, holds real conversations, captures outcomes, and learns from every call — fully autonomous, zero human touch for standard calls.

Deployed for FundingAssist.com.au — Melbourne Mortgage Brokerage

Explore Architecture Work Together

Live Dashboard

Workflow Depth

Decision tree levels

Daily Capacity

100+

Autonomous calls per day

AI Assistants

Specialized VAPI squad agents

Resolution Rate

85%+

First-call success target

The Problem

Outbound Calling at Scale
Breaks Without Automation

Manual outbound calling is one of the most expensive and inconsistent operations in any service business. Staff get stuck in IVR loops, miss callbacks, forget to log outcomes, and can't handle volume. Brokers at FundingAssist.com.au were spending hours each day on routine follow-up calls that yielded inconsistent results.

The challenge wasn't just volume — it was the complexity of real-world telephony. Calls hit IVR menus, go to voicemail, transfer multiple times, and require structured conversations with specific outcomes logged. A simple auto-dialer solves none of this.

What was needed was a system that could handle the entire call lifecycle autonomously — navigating phone trees, identifying who answered, conducting a real conversation, capturing structured data, scheduling callbacks, and escalating to a human only when genuinely necessary.

The platform needed to be self-improving, learning the patterns of each contact so every subsequent call was faster and smarter than the last.

Before the Platform

Manual outbound calling pain points

Hours of broker time lost to routine follow-up calls daily
IVR menus caused agents to abandon calls or navigate incorrectly
No consistent call outcome logging — data scattered across emails and notes
Callbacks missed when staff unavailable — leads going cold
No retry logic — a busy signal meant the lead wasn't followed up
Zero visibility into call performance, resolution rates, or escalation patterns
No learning between calls — same IVR navigated blindly every time

System Architecture

18-Level Autonomous
Call Workflow Engine

Built end-to-end in CodeIgniter 3 PHP — a modular, production-grade system where each workflow phase handles a distinct real-world telephony scenario with full logging, learning, and fallback logic.

Live Production System

Five-Phase Autonomous
Call Lifecycle

From trigger to outcome, every call passes through five distinct phases — each with its own AI assistant, decision logic, and failure recovery path. Human intervention triggered only when AI confidence falls below threshold.

Phase 1 — Initiation & Validation

Rate limit check, readiness validation, duplicate call detection, E.164 phone formatting, VAPI squad call trigger with dynamic system prompt injection.

Phase 2 — Connection Detection

Real-time classification: human answered vs. IVR detected vs. voicemail vs. no-answer/busy. Each path routes to a dedicated specialist AI assistant optimized for that scenario.

Phase 3 — IVR Navigation

Intelligent menu traversal using stored patterns from the Clinic Intelligence database. On first call, the AI learns and records the IVR path. On every subsequent call, it replays the known winning sequence.

Phase 4 — Live Conversation

GPT-4 powered structured conversation with staff identification, real-time confidence scoring, outcome capture, and decision logging. Controlled substance and compliance flags built into conversation logic.

Phase 5 — Resolution & Follow-Up

Outcome recorded, callback scheduled if needed, escalation queue updated for human intervention cases, alternative contact (fax/email) triggered on voice failure. Full audit trail written to 12+ database tables.

Call Data Flow

VAPI Squad Trigger

10 specialist assistants · dynamic prompt injection

Deepgram Nova-2 Transcription

Real-time speech-to-text · every call word captured

GPT-4 Conversation Engine

Confidence scoring · structured outcome extraction

MySQL — 12+ Workflow Tables

Call logs · IVR patterns · intelligence · analytics

Clinic Intelligence Engine

Self-updating · learns IVR paths, staff, best times

Escalation Queue → Human

Low-confidence triggers · full context passed over

18Levels

10Agents

100+Calls/Day

Key Features

What the Platform Does

Every feature is production-built and battle-tested against real-world telephony — not a prototype or proof of concept.

Core Engine

18-Level Decision Workflow

A multi-tier routing engine that covers every real-world call scenario — from initial answer detection through nested IVR menus, staff transfers, voicemail, and escalation paths.

Human vs. IVR vs. voicemail detection at connection
Smart routing to specialist assistant per scenario
Stuck-workflow detection with automatic recovery
Full decision log written at every branch point

IVR Intelligence

IVR Pattern Memory & Navigation

The system detects automated phone menus, navigates them using stored knowledge, and records the winning key-press sequence for every new IVR encountered — improving with each call.

First-call IVR mapping — patterns stored to database
Learned sequences replayed on subsequent calls
Immediate termination on unresolvable IVR loops
Simple, multi-level, and voice-activated IVR handled

AI Layer

VAPI Squads + GPT-4 Engine

Ten specialized AI assistants orchestrated by a VAPI Squad controller — each fine-tuned for a specific call scenario with domain-specific system prompts and confidence thresholds.

10 assistants: greeting, IVR, human conversation, voicemail, escalation and more
GPT-4 with controlled temperature and token budgets
Real-time AI confidence scoring per response
Deepgram Nova-2 for accurate live transcription

Self-Learning

Clinic Intelligence Engine

A continuously updated knowledge base that captures everything learned from each call — IVR structure, staff names, authorization levels, best call times, and outcome patterns.

Stored per contact: IVR path, staff, best time, VCPR
Intelligence loaded at call start for instant advantage
Staff authorization levels tracked per contact
System measurably improves after first 100 calls

Reliability

Smart Retry & Escalation Queue

When calls fail, the system retries intelligently — with rate limiting, exponential logic, and callback scheduling. When AI confidence drops, the call is escalated with full context preserved.

Exponential retry backoff with configurable limits
Callback scheduling with timezone-aware timing
Escalation queue with priority scoring by reason
Human operators receive full call context on handoff

Analytics

Real-Time Dashboard & Health Monitor

A full analytics layer tracking success rates, AI confidence scores, escalation rates, call durations, IVR navigation success, and callback completion — with a live health check endpoint.

30+ controller endpoints for monitoring and management
Workflow-level distribution and bottleneck analysis
Health check validates DB, API, and config at any time
Batch processing for high-volume call campaigns

Tech Stack

Built on Production-Grade Infrastructure

Every component chosen for reliability, performance, and cost efficiency in a real-world telephony environment.

PHP 7+

CodeIgniter 3

VAPI Voice AI

OpenAI GPT-4

Deepgram Nova-2

Twilio

MySQL

VAPI Squads Architecture

RESTful Webhooks

E.164 Phone Formatting

PHP 7+

CodeIgniter 3

VAPI Voice AI

OpenAI GPT-4

Deepgram Nova-2

Twilio

MySQL

VAPI Squads Architecture

RESTful Webhooks

E.164 Phone Formatting

Transient Assistants

Cron Job Automation

Smart Retry Logic

Callback Scheduler

JSON API

IVR Pattern Engine

analysisPlan Schema

Rate Limiting

Escalation Queue

Health Monitoring

Transient Assistants

Cron Job Automation

Smart Retry Logic

Callback Scheduler

JSON API

IVR Pattern Engine

analysisPlan Schema

Rate Limiting

Escalation Queue

Health Monitoring

Client Spotlight

Deployed for FundingAssist.com.au

Live Deployment · Melbourne, AU

FundingAssist.com.au
Mortgage Brokerage

FundingAssist is a Melbourne-based mortgage brokerage helping Australian home buyers and investors secure the right loan. With a high volume of inbound leads requiring timely follow-up, the brokerage deployed the AI Calling Agent Platform to automate outbound lead follow-up, applicant qualification calls, and documentation status checks — freeing brokers to focus on high-value advisory work rather than routine outbound dialling.

Visit FundingAssist.com.au

Lead Follow-Up

Automated outbound calls to new loan enquiries

Lead Qualification

AI conducts structured qualification conversations

Document Status Checks

Automated calls to chase outstanding application docs

Broker Time Freed

Routine calls handled autonomously — brokers focus on advice

Notable Achievements

What Makes This Platform Stand Out

Built solo — every line of architecture, code, and integration designed and delivered by one engineer.

18 Workflow Levels

Covers every real-world telephony scenario — IVR, voicemail, transfers, authorization, compliance, and more

10 Specialist AI Agents

VAPI Squads architecture — each assistant fine-tuned for one call scenario with its own system prompt and thresholds

3 AI Providers Integrated

VAPI orchestration · OpenAI GPT-4 language model · Deepgram Nova-2 transcription — all in one coherent system

85%+ Resolution Rate Target

First-call resolution goal — achieved after the system's 100-call learning phase with each new contact group

8k+ Lines of Production Code

PHP controllers, models, SQL schema, and frontend — built solo, end-to-end, with full documentation and test coverage

Self-Learning Intelligence Engine

IVR patterns, staff profiles, and call timing data stored and improved with every completed call — compounding intelligence

My Role

Built Solo — End to End

Every layer of this platform — from architecture decisions to production deployment — was designed and delivered by one engineer.

System Architect

Designed the 18-level decision workflow, VAPI Squads structure, 10-assistant configuration, and the self-learning intelligence engine from first principles.

Backend Developer

Built all PHP controllers, models, and API endpoints in CodeIgniter 3 — 8,000+ lines of production code including 30+ controller endpoints and 50+ model functions.

AI Integration Engineer

Integrated VAPI, OpenAI GPT-4, and Deepgram Nova-2 — including the migration from VAPI's deprecated Workflows to the modern Squads + Assistants architecture.

Database Designer

Designed and implemented 12+ MySQL tables for call tracking, IVR pattern storage, clinic intelligence, escalation queuing, callback scheduling, and analytics.

DevOps & Reliability Engineer

Configured webhooks, cron jobs, rate limiting, health monitoring endpoints, stuck-workflow detection, and multi-layer error recovery for production reliability.

Want an AI Calling Agent for Your Business?

Whether you're running a mortgage brokerage, an e-commerce operation, a real estate sales team, or a service business — if outbound calling is costing you time, we can automate it. Let's talk.

hello@singhaisystems.comRespond within 24 hours

singhaisystems.comView full portfolio and services

LinkedInConnect with Prabhdeep Singh

Send a Message

✓ Message sent! I'll be in touch within 24 hours.

Your Name

Email Address

Company / Website

I'm interested in

Tell me about your outbound calling needs