Respondi

Overview

A multi-tenant platform for building AI agents connected to real messaging channels: WhatsApp through Meta's official Cloud API, and Telegram. Each agent has its own knowledge base with RAG retrieval, connectors to external tools, and a queue-backed engine that keeps conversations coherent. It generalizes the parts of Gastronova's agents that were worth reusing — Gastronova is its first intended consumer.

The Challenge

Most WhatsApp bots are built on unofficial bridges that break without warning and can get the number banned. Doing it properly means the official Cloud API, strict webhook signature validation, and a design where every tenant's channel credentials are isolated and encrypted.

The Approach

Split the system into a Fastify core API (agent engine, RAG, background workers), a thin channel gateway that only validates and normalizes provider webhooks, and a Next.js front end — with Zod contracts shared between them. Schema changes go through Drizzle migrations that run before the containers come up; media lives in a private S3-compatible bucket.

The Outcome

Deployed and running behind its own domain since 2026, with the full container stack, migrations wired into the deploy, and channel onboarding in place. Next up: connecting the first production WhatsApp Business account.

Key Highlights