The Problem

Millions of small and mid-sized businesses — especially across Africa and emerging markets — run their entire customer support on WhatsApp. Yet the tooling available is either enterprise-priced, dumb (no AI), or introduces unbearable latency for distributed users.

I wanted to build something that's:

AI-native from day one (not bolted-on)
Multi-region with sub-100ms database latency globally
Accessible to any business with a WhatsApp number
No-code friendly for building complex chatbot flows

What I Built

ImaradeskAI — a full-stack WhatsApp customer support platform that turns any WhatsApp Business number into an intelligent helpdesk with AI-powered automation, visual chatbot flows, and real-time analytics.

🧠 Multi-Model AI Engine

Three providers working together — Gemini 2.5 Flash, GPT-4o, and Claude Sonnet 4 — switching transparently based on task type and availability.

Every inbound message gets:

Intent classification into 9 categories with confidence scoring
Sentiment analysis (-1 to +1) feeding escalation risk
Language detection + auto-translation — respond in the customer's native tongue
Conversation summarization for automatic ticket generation

The system caches classifications and embeddings, so repeated queries are near-instant.

🤖 Visual No-Code Chatbot Builder (13 Node Types)

A drag-and-drop flow canvas where non-technical users can build complex conversation logic without writing code:

Node Type	What It Does
Trigger	Entry point — keyword, regex, business hours, first message
Send Message	Send text to the customer
Quick Replies	Present numbered options for structured responses
AI Response	Single-turn AI generation using conversation context
Agent (Full AI)	Multi-turn autonomous agent with memory + KB access
Knowledge Base	Vector similarity search + AI-synthesized answer
Condition	Multi-rule branching (contains, regex, equals, not_equals)
Delay	Configurable wait (optimized for serverless at ≤5s)
Connector	Merge multiple branches into one path
Action	Side effects: create ticket, assign agent, add tag, close
API Call	External HTTP calls with SSRF protection (blocks private IPs)
Smart Reply	Template-based replies with shortcut matching
Human Handoff	Escalate to agent + auto-create ticket + assign team

The flow engine supports:

Trigger priority matching (specific keywords before catch-all)
Depth-limited execution (max 20 nodes) to prevent infinite loops
Stateful conversations — activeFlowId + currentNodeId per contact
Full execution logging for debugging and analytics

📚 RAG Knowledge Base with Vector Embeddings

Upload (PDF/DOCX/TXT/CSV) → Extract Text → Embed → Store → Cosine Search → AI Generate

Enter fullscreen mode Exit fullscreen mode

Documents embedded with gemini-embedding-001 or text-embedding-3-small
Vectors stored as JSON arrays directly in Aurora DSQL — no separate vector DB needed
Cosine similarity ranks all KB entries against the user's query
Top-k results passed as context to AI for grounded, hallucination-resistant answers
Files managed in AWS S3 with UUID-based naming for security

🌍 Multi-Region Database with AWS Aurora DSQL

This was the biggest architectural bet. Aurora DSQL gives us:

Two active regions (us-east-1 + us-east-2) with automatic geo-routing
Short-lived IAM tokens via DSQL Signer (auto-refreshed every 10 min)
Connection pooling per region with failover to secondary
Client lat/lon routing — the nearest cluster handles each request

function getNearestRegion(lat: number, lon: number) {
  const regions = [
    { endpoint: US_EAST_1_ENDPOINT, lat: 39.0, lon: -77.5 },
    { endpoint: US_EAST_2_ENDPOINT, lat: 40.0, lon: -83.0 },
  ];
  // Euclidean distance → return closest endpoint
}

Enter fullscreen mode Exit fullscreen mode

The result: sub-100ms database responses regardless of where the user is connecting from.

🎫 Intelligent Ticketing + SLA Engine

Tickets are auto-created from AI conversation analysis with generated titles and descriptions. SLA policies define per-priority timers, and an escalation risk score (0-100) considers:

Factor	Points
SLA due within 30 min	+40
Urgent priority	+25
Customer frustration (4+ follow-ups)	+15
Negative sentiment streak	+15
Stale (8+ hrs no update)	+10

📊 AI-Powered Analytics

AI Copilot — suggests replies to agents with acceptance rate tracking
Voice of Customer — sentiment distribution, complaint detection, praise identification
Agent Quality Score — resolution rate (40%) + sentiment impact (35%) + activity (25%)
Contact Health Score (0-100) — engagement recency, open tickets, message frequency
Topic Trending — intent volume week-over-week with severity classification
Workload Distribution — agent capacity management with concurrent chat limits

📢 Broadcast Campaigns

Scheduled or immediate mass messaging via WhatsApp templates
Audience segmentation with JSON-based filters (tags, segments, contact lists)
Real-time delivery analytics: sent → delivered → read → replied → failed
Template variable substitution: {name}, {phone}, custom fields

Tech Stack

Layer	Technology
Framework	Next.js 16, React 19, TypeScript
Styling	Tailwind CSS 4, Radix UI, shadcn/ui
Database	AWS Aurora DSQL (PostgreSQL-compatible, multi-region)
ORM	Drizzle ORM
File Storage	AWS S3
AI Models	Gemini 2.5 Flash, GPT-4o, Claude Sonnet 4
Embeddings	gemini-embedding-001, text-embedding-3-small
Auth	JWT (jose) + bcrypt, httpOnly cookies
Messaging	Meta WhatsApp Cloud API (Graph v21.0)
Deployment	Vercel (Edge + Serverless)
Charts	Recharts
Forms	React Hook Form + Zod

What's Next

More regions (eu-west-1, af-south-1 for African markets)
Voice message transcription with Whisper
CSAT surveys sent via WhatsApp after ticket resolution
Shopify/WooCommerce integration for order status automation
Team performance leaderboards with gamification

Live Demo: imaradesk-ai.vercel.app

If you're building WhatsApp-based tools or working with Aurora DSQL, I'd love to hear your experience. Drop a comment below!