Skip to content
Back to Projects
ProductionAI/MLBackendFull-StackCo-Founded

Zaylon AI

Multi-agent conversational commerce system built with LangGraph, serving real users across MENA markets via WhatsApp and Instagram.

LangGraphPythonMulti-AgentShopifyNext.js

The Problem

MENA merchants lose sales because customers abandon messaging channels when they can't get instant, dialect-aware product assistance. Traditional chatbots fail with Arabic dialects and can't handle complex multi-step purchases.

Architecture & Approach

Built a multi-agent supervisor architecture with LangGraph that routes conversations to specialized agents, each with scoped tool access and dialect-aware NLP. The system handles the full purchase lifecycle from product discovery to payment processing across multiple messaging channels.

Key Technical Decisions

LangGraph over plain LangChain

State machine routing gives deterministic conversation flow control that chain-based approaches can't guarantee. Each conversation phase (browsing, carting, checkout) has well-defined transitions.

Supervisor pattern over flat multi-agent

Scoped tool access per agent prevents hallucinated actions and reduces the attack surface. A sales agent should never be able to process payments.

Redis message accumulation

WhatsApp users send 3-5 rapid messages instead of one coherent prompt. Without batching, each triggers a separate agent invocation, wasting tokens and producing fragmented responses.

Results

Production system serving merchants across 6 e-commerce platforms, handling tri-lingual conversations with automated sales, support, and checkout flows.