AI WhatsApp CRM & Voice Receptionist: Autonomous Lead Qualification
A multi-tenant conversational CRM where AI agents process incoming WhatsApp texts and Darija/Arabic/French voice notes, answer catalog questions, qualify leads, and hand off hot deals to human reps.
LaravelNext.jsWhatsApp AIPostgreSQL
Context
Standard CRMs assume reps manually log emails and web forms. Across Morocco and MENA, sales happen over WhatsApp messages and voice notes mixing Darija, Arabic, and French — and leads go cold waiting hours for a manual reply.
Scope I disclose publicly
- Webhook ingestion backed by Redis queues so bursts of incoming WhatsApp texts, voice notes, and images are processed asynchronously without dropped events.
- Dialect-aware speech-to-text pipeline handling Moroccan Darija, Arabic, and French audio before passing context to the reasoning agent.
- Tool-calling agent layer grounded against live product catalogs and calendar availability so pricing and booking replies never hallucinate.
- Human-in-the-loop takeover that pauses automation when a human rep steps in or when high-intent escalation rules fire, complete with a conversation summary.
Decisions and trade-offs
- Decoupling webhook acknowledgement from audio transcription and LLM execution via Redis queues to guarantee sub-second webhook responses.
- Restricting agent answers to verified catalog and policy tools rather than open-ended generation.
- The trade-off: voice-note transcription plus tool-calling adds a few seconds of latency and per-message API cost compared to rigid button menus.
This is a private multi-tenant codebase; live sandbox demonstrations of the WhatsApp and voice pipeline are available on request.
Disclosure: No customer phone numbers, message logs, or client conversion metrics are published.