I built this to manage conversations across WhatsApp, Telegram and Instagram from one backend. Incoming messages can be routed to different agents, each with its own instructions, model settings and knowledge base. I deployed and used it previously; it is no longer running.
I made the channel accounts, routing rules and agent settings configurable in the admin area. That let me change who handled a conversation, which knowledge base they used and when a human should take over without rewriting the webhook handlers. The operational screens also made it possible to inspect failed responses and model usage alongside the conversations.
The useful part was being able to see the whole conversation and intervene when needed. I added an inbox, human handoffs, knowledge retrieval and model usage tracking, along with a fallback model when the primary one failed. The FastAPI backend and administration screens are in the GitHub repository. This work also fed into how I approached messaging in the meditation portal.
