Conversa AI — Multi-Tenant SaaS for Conversational AI Agents

Overview

Conversa AI is a multi-tenant SaaS for conversational AI agents focused on automated support and commercial conversion. The goal was to move from Level-2 chatbots to Level-3 AI agents with tool execution—assistants that can query knowledge bases, invoke tools, and run conversations with real context.

Beyond a unified inbox, the project is a full systems engineering exercise: a message pipeline, LLM orchestration, RAG with vector embeddings, a rules engine, real-time sentiment analysis, and omnichannel routing behind one Message Gateway.

Technical architecture & ownership

• Full-stack engineering: End-to-end TypeScript with Next.js 16 (App Router + React 19 + React Compiler) on the frontend and NestJS 11 on the backend.
• Real multi-tenancy: All data scoped by organizationId via TenantInterceptor and TenantGuard, with plan-based feature gating through @RequireFeature().
• AI & RAG: OpenAI and Anthropic SDKs, 1536-dimensional embeddings stored in PostgreSQL + pgvector, and an Agent Engine that registers tools dynamically via OnModuleInit.
• Infra & DevOps: Docker Compose in production, CI/CD with GitHub Actions (buildx → GHCR → SSH deploy), BullMQ for async processing, and MinIO (S3-compatible) for media storage.
• Real time: Socket.IO via @nestjs/websockets for a live inbox, typing indicators, and sentiment updates as conversations progress.

Stack

• Frontend: Next.js 16, React 19, Tailwind CSS v4, Zod v4, next-intl (PT/EN)
• Backend: NestJS 11, Prisma ORM (~60 models), PostgreSQL 17 + pgvector, Redis 7
• AI: OpenAI SDK, Anthropic SDK, RAG with vector embeddings
• Integrations: Evolution API (WhatsApp), Meta Graph API (Instagram/Messenger), SendGrid (email), n8n (external automations)
• Infra: Docker Compose, MinIO, BullMQ, Socket.IO, JWT auth

Supported channels

Five channels integrated into the same Message Gateway:

• WhatsApp via Evolution API
• Embeddable Web Chat via Socket.IO widget
• Email via SendGrid (inbound + outbound)
• Instagram DM via Meta Graph API
• Facebook Messenger via Meta Graph API

Message pipeline

Every inbound message goes through a deterministic chain:

1. Identification — create or resolve contact and conversation
2. Auto-assign — round-robin or least-busy across department agents
3. Persistence + async analysis — sentiment and language detection in the background
4. Rules engine — evaluate tenant-configured automations
5. Chatbot flows — try to match a structured flow
6. AI agent fallback — if nothing matches, the LLM with RAG takes the conversation

Engineering highlights

• AI Copilot suggesting three replies for the human agent—faster handling without blind automation.
• Conversation summary and sentiment analysis in real time, updating the UI as the chat evolves.
• Per-tenant rules engine (SLA, priority, tags, assignment) without redeploys.
• Per-tenant rate limiting (10 req/s, 50 req/10s, 200 req/min) via a global TenantThrottlerGuard.
• Operations-first: health checks, soft delete, activity log, contact merge, and CSV exports.

Next steps

• Visual drag-and-drop flow builder (today: JSON editor)
• 2FA/SSO and a public API documented in Swagger
• Automated test coverage and structured logging
• Telegram Bot API channel