AI BeeOmnichannel AI Sales & Support Agent

It is a Turborepo monorepo: a Next.js 16 / React 19 dashboard and marketing site, an Express + MongoDB API, a React Native companion app for the team, and a Shopify extension.

How one inbound message is handled:

1. Meta sends a webhook; the API verifies its `X-Hub-Signature-256` signature and deduplicates it by message id.
2. The turn is queued on BullMQ + Redis, and rapid-fire messages from the same person are buffered into a single turn.
3. Voice notes are transcribed and images described with Gemini; the conversation language is detected.
4. A RAG step retrieves the workspace's own knowledge base (Gemini embeddings, optional cross-encoder reranking) and a semantic cache can serve a known answer without an LLM call.
5. The LLM agent (@openai/agents over DeepSeek) writes the reply with tool calls for meeting scheduling; persona, lookup and groundedness guards check it before it is sent.
6. An action router creates meetings, notifies the team or hands the thread to a human, and emits events into a rules-based automation engine.

Around that pipeline the product includes:

• Lead CRM with AI lead scoring, AI conversation summaries and bulk actions.
• Comment moderation and per-post comment controls.
• Human takeover: a teammate can pause the AI on one thread and hand it back.
• Built-in video meetings (Agora) with booking from the chat.
• Post scheduling, AI content and blog generation.
• Stripe subscriptions metered by token usage, a partner programme, and a super-admin console.
• Role-based access per workspace, an audit trail, and access tokens encrypted at rest.
• Push notifications to the mobile app via Firebase.

1. Replies that stay grounded in the business's own facts: A sales agent that invents prices or delivery times is worse than no agent. I built a RAG pipeline over each workspace's knowledge base and put guards after the LLM (persona, lookup and groundedness checks, plus a final sanitiser) so a reply that isn't supported by the knowledge base is retried or rewritten before the customer sees it.
2. Never promising what the team won't deliver: When the AI tells a customer "our team will call you", that has to be true. Replies that commit the team are gated on a push alert actually reaching a teammate first; if the alert fails, the reply is rewritten so it doesn't make the promise.
3. Webhooks at real-world volume: Meta retries deliveries and customers send five short messages in a row. Signature verification, idempotent event claims, a BullMQ job queue, per-conversation turn locks and a message-fragment buffer keep each customer turn answered once, in order, with full context.
4. Handing a conversation between AI and humans cleanly: Teams need to jump into a thread without the bot talking over them. I implemented human takeover across the messaging channels, so the AI stays quiet while a teammate is in the thread and every message shows whether the AI or a person sent it.
5. Keeping LLM cost and latency down: A semantic reply cache, optional reranking with a tight timeout, and model experiments behind flags keep replies fast while subscriptions stay metered by token usage.