SirenAI · sirenai.net

Case Study · Multi-Tenant SaaS / AI Agents

SirenAI: From Manual Texting to Autonomous Bookings

Multi-channel AI messaging & booking platform.

How we rebuilt a stalled SaaS into a 24/7 AI booking concierge across WhatsApp, SMS, and Telegram, with a self-hosted backend, a real unified inbox, and an agent that handles tier-1 conversations end-to-end.

Modern messaging interface, representative imagery for the SirenAI multi-channel AI platform

~24 hr → <3 sec

Off-hours reply latency

Tier-1 → AI

100% operator workload offloaded

0 → 3

Channels in a unified inbox

Zero

Double-bookings after rebuild

ScopeArchitecture · Backend · AI Agent · Multi-channel Messaging · Admin Dashboard · DevOps

01About the client

Industry: Multi-tenant SaaS for boutique service agencies.

Scale: Hosts dozens of agencies, hundreds of providers, thousands of clients per tenant.

Context: Operators manage bookings, client verification, payments, and cancellations across WhatsApp, SMS, and Telegram simultaneously.

Constraints: Strict tenant data isolation, WhatsApp/SMS compliance, data sovereignty preferences.

02The problem

A stalled SaaS, paying a compounding tax.

The legacy platform created compounding inefficiencies. Manual texting across three channels. Operators handling 100% of conversations because the prior vendor's AI never shipped. WhatsApp and SMS numbers getting banned. App-layer tenant isolation that nobody trusted.

Operator manually messaging across multiple phones, the kind of workflow the rebuild replaced

Pain point	Impact
Manual texting across 3 channels	~24-hour reply latency; lost bookings
Incomplete AI from previous vendor	Operators handled 100% of conversations
Broken multi-tenant isolation	Risk of data leakage between agencies
WhatsApp/SMS numbers getting banned	Channel outages; lost client trust
Unpredictable per-message LLM costs	Budget volatility
No unified inbox	Operators juggling 3 apps per shift
No booking conflict prevention	Double-bookings; refund disputes

03Solution

Four components. One vertically integrated platform.

Admin Dashboard

React 19 · Refine v5 · shadcn/ui · Tailwind

Unified inbox, booking calendar, provider management, client verification, payments, role-based dashboards, and a cross-tenant admin console.

Messaging Pipeline

Express 5 · BullMQ · Redis

One webhook per channel, unified job queue, structured logging with correlation IDs. Worker concurrency 5 with retries and exponential backoff.

AI Agent

Mastra · Vercel AI SDK · pgvector

Postgres-backed memory. Handles bookings, verification, and handoffs autonomously. Persona trained from historical messages. Dynamic system prompts based on tenant policy, language, and rules.

Shared Backend

Self-hosted Supabase · PostgreSQL 15 · Realtime · Storage

RLS policies enforce tenant isolation at the database layer. Atomic operations via SECURITY DEFINER RPCs guarantee multi-table consistency.

04Implementation

Phased milestone delivery with client review at each stage.

01
Backend foundation
Multi-tenant platform ready for AI and messaging.
02
Messaging dashboard
Real-time unified inbox with manual operator fallback.
03
AI conversation agent
Handles bookings and client verification autonomously.
04
Booking engine
Conflict prevention and calendar management.
05
SMS & multi-channel routing
Seamless multi-channel communication and rapid tenant onboarding.

Challenges overcome

Rapid tenant onboarding

Modified the SMS gateway and onboarding flow, customizing server and client packages so spinning up a new tenant doesn't need engineering work.

AI persona that doesn't sound like a bot

Distilled from real historical operator messages, then layered with real-time persona-leak scoring to catch drift before it goes out.

Sub-3-second responses

Optimized memory writes, trimmed context windows, and added typing indicators so the conversation feels natural even when the model is thinking.

05Technical decisions

Decisions that earned their footprint.

Every choice traded clearly: when we picked Postgres exclusion constraints over application-level locks, we wanted concurrency safety the DB itself guarantees. When we picked a provider-agnostic LLM layer, we wanted to swap models in a weekend if pricing or capability shifted.

Tenant Isolation

Database-enforced via RLS policies, not app-layer trust.

Atomic Operations

Multi-table RPCs guarantee consistency across booking, billing, and conversation state.

Booking Conflict Prevention

Postgres exclusion constraints. The DB rejects overlaps, no race condition possible.

Channel Abstraction

Single message-job structure for all three channels; new channels need a webhook + dispatcher, nothing else.

AI Memory

Stored in Postgres with pgvector. No separate vector DB to operate.

Guardrails

Prompt injection detection, PII scrubbing, and persona-leak scoring on every model call.

06Integration footprint

Five integrations, each chosen for a reason.

Integration	Mechanism	Reason
WhatsApp	Self-hosted gateway	High-density numbers; no Meta dependency.
SMS	Android APK + gateway	Rapid onboarding; tenant-controlled SIMs.
Telegram	Bot API	Free, HTTPS webhook only.
Payments	BTCPay Server	Non-custodial, open-source.
LLM	Mastra + Vercel AI SDK	Provider-agnostic. Swap models without code changes.

Scalability

Stateless workers, shared Redis and Postgres for load distribution.

Testing

400+ automated tests ensure reliability across dashboard and pipeline.

07Business impact

Before, and after.

Modern SaaS dashboard view, representative imagery of the unified admin platform

Strategic gains

24/7 coverage without a night shift
Self-service tenant onboarding
Channel-ban resilience across WhatsApp, SMS, and Telegram
Vendor-portable, open-source stack

Metric	Before	After
Off-hours reply latency	~24 hr	<3 sec
Unified inbox channels	0	3
Manual operator workload	100%	AI handles tier-1
Double-bookings	Recurring	Zero
Tenant data isolation	App-layer only	DB-enforced
Multi-language support	Broken	LLM-native
Time-to-add-channel	Vendor-blocked	One webhook + dispatcher

08Client testimonial

“Unlike the previous vendor, which struggled for months, IDEfforts focused on finding real solutions and delivering results. Within weeks, messages were being answered and bookings handled autonomously. The platform finally works the way we needed.”

Founder · SaaS booking platform (anonymized)

09Key takeaways

What we'd carry into the next build.

01
RLS in multi-tenant SaaS eliminates entire classes of vulnerabilities.
02
Provider-agnostic LLMs let you swap models without rewriting application code.
03
Mining real conversations beats prompt-engineering for persona quality.
04
DB-enforced booking constraints prevent concurrency bugs no application code can.
05
One queue for multiple channels reduces duplicated logic to near zero.

10Tech stack at a glance

What it runs on.

Layer	Technology
Frontend	React 19 · Refine v5 (headless) · TypeScript · Vite
UI	shadcn/ui · Tailwind CSS · Radix
Backend	Supabase (self-hosted) · PostgreSQL 15 · Realtime · Storage · pgvector
Pipeline	Node.js · Express 5 · BullMQ · Redis
AI	Mastra Agent · Vercel AI SDK · provider-agnostic LLM
Channels	WhatsApp gateway · SMS gateway · Telegram Bot API
Payments	BTCPay Server
Testing	Vitest · 400+ integration tests

Practice

Got a stalled SaaS or AI roadmap that won't ship?

We deliver production-ready AI platforms in weeks, not quarters. Tell us about your stack. We'll tell you honestly whether we're the right fit.

Start a conversation