Hybrid AI CRM Chatbot

A sales agent for Messenger and Instagram DMs that runs authored conversation flows, not generated text. The language model classifies what the customer means; every reply is pre-written in YAML — so responses stay predictable, on-brand, and controllable per client. One deployment serves many pages, each with its own script and knowledge base.

Enquire about a deployment
Messenger Instagram DM FastAPI Multi-Page

What it does

The chatbot receives webhook events from Meta Messenger and Instagram DMs, works out what the user wants, and replies with scripted conversation flows defined in YAML. The conversation logic is a deterministic state machine — branching, objection handling, escalation to a human — with the language model used at one specific point: scoring which authored branch best matches what the customer just said.

Because responses are pre-written rather than generated, they never hallucinate a price, invent a policy, or drift off-brand. The same engine works for a car dealership, a dental clinic, or a beauty salon without code changes — each vertical is just a different script, knowledge base, and set of classifier hints.

Authored Flows Conversations are YAML trees of branches — each with an intent description, keyword fallbacks, multilingual responses, and transitions. Human-editable, version-controllable, no database required. Every reply the customer sees was written and approved in advance.
LLM for Intent Only A multi-provider classifier (OpenAI → Gemini → OpenRouter, with automatic failover) determines the user's intent, urgency, and pain points using page-specific knowledge base context. It selects a branch; it never writes the message.
Multi-Page Mode A single deployment serves multiple Meta Pages, each with its own script, knowledge base, access token, and routing thresholds. Incoming webhooks route by Page ID; conversation state is isolated per page. New clients onboard onto the shared platform without rebuilding the integration.
Persistent State Conversation state lives in Redis with a configurable TTL (30 days by default), keyed per page and user. It survives restarts and supports re-entry — resuming a conversation naturally after a customer goes quiet and comes back.
Platform-Agnostic Messenger and Instagram payloads are normalised into one internal message model before processing, so the rest of the pipeline doesn't care which platform a message came from. Rapid-fire messages are batched within a short window so the bot replies to the whole thought, not each fragment.
Integrations Optional Google Calendar booking for appointments, CRM follow-up scheduling, live inventory polling with fuzzy search, and fire-and-forget Postgres analytics that never block the reply pipeline. Each is opt-in per deployment.

Why hybrid, not generative

A fully generative chatbot is fluent but unaccountable: it will confidently state a price you never set or a policy you don't have. A pure keyword bot is safe but brittle, breaking the moment a customer phrases something unexpectedly.

This is the middle path. The LLM handles the one thing it's genuinely good at — understanding messy, multilingual, real-world phrasing — and hands off to authored copy for everything the customer actually reads. You get the naturalness of an LLM front door with the control of a scripted back end, and you can audit every possible response by reading the YAML.

Frequently asked questions

Does the bot generate its own replies?

No. Every response is pre-written in your conversation script. The language model only classifies intent — it chooses which authored branch to send. This is what keeps responses predictable and on-brand.

Which channels are supported?

Meta Messenger and Instagram DMs. Both are normalised into one internal message format, so a single script serves both channels.

Can it handle multiple languages?

Yes. Responses are authored per language in the script, with heuristic language detection and a configurable fallback. Intent classification is language-agnostic.

Can it serve more than one business?

Yes. In multi-page mode, one deployment serves many Meta Pages, each with its own script, knowledge base, and access token. Adding a client is a config change, not a rebuild.

What happens if a customer needs a human?

Escalation is a first-class branch in the state machine. Scripts define when and how to hand off, and CRM follow-ups can be scheduled to re-engage a lead later.

Is it a product I can buy?

It's deployed as a managed service. Quiet Signals Lab runs the platform and authors the conversation script and knowledge base for your business. Get in touch to discuss a deployment.

Get in touch

The Hybrid AI CRM Chatbot is built and operated by Quiet Signals Lab, and already deployed for a social media marketing agency's client accounts. If you'd like a bot for your business, or want to onboard a set of client pages onto the platform, reach out.

Tell us the channels you use, the languages you serve, and what a good conversation looks like for your business — and we'll take it from there.

hello@quietsignalslab.com →