Zaylon AI

The problem

MENA shoppers do a lot of their buying over WhatsApp and Instagram, but merchants can't sit in those chats around the clock. Off-the-shelf chatbots fall apart on Egyptian Arabic and Franco-Arabic, and they can't carry someone from a question all the way to a paid order.

Three dialects, one conversation

Egyptian customers rarely write in Modern Standard Arabic. They mix dialect, Franco-Arabic, and English, sometimes in one sentence. Zaylon detects the dominant language on every turn and replies in the same register, so the same question can arrive three ways and get the same answer back in the way it was asked.

English

Hi, is the M size back in stock?

Egyptian Arabic

السلام عليكم، مقاس الميديوم متوفر؟

Franco-Arabic

salam, el medium raga3 el stock?

Approach

I built it in LangGraph as one tool-calling agent with about 35 scoped tools spanning sales, support, and checkout. It began as a multi-agent supervisor that routed to separate Sales, Support, and Checkout agents; I consolidated it into a single agent that proved more reliable and easier to reason about. Dialect detection runs first, so the customer is answered in the language they wrote in, and the same flow carries a conversation from the first product question to a confirmed payment.

Decisions

1. 01

   LangGraph over plain LangChain

   A state machine lets me pin down exactly which transitions are legal between browsing, carting, and checkout. Plain chains drift, and in a flow that ends in a payment I can't afford drift.

2. 02

   Consolidated to a single tool-calling agent

   It started as a multi-agent supervisor routing to separate agents. A single agent with scoped tools turned out to be more reliable and far easier to reason about: fewer handoffs to get wrong and one place to trace a conversation, while tool-level scoping still limits what any single step can touch.

3. 03

   Redis message accumulation

   People on WhatsApp send three or four quick messages instead of one. I batch them in Redis for a short window so the agent reads the whole thought at once, instead of firing on each fragment and answering three times.

What it does

• Handles the whole conversation end to end: answers product questions, builds the cart, and closes the order, all inside the same chat and without a human stepping in.
• DialectBridge detects whether a customer is writing English, Egyptian Arabic, or Franco-Arabic and answers back in the same dialect.
• Customers can send a photo of a product, and Zaylon matches it against the merchant's catalog.
• Follows up on abandoned carts and steps in when a conversation starts to turn negative.
• A merchant dashboard with a live inbox, analytics, RFM segmentation, and A/B testing.
• Each merchant's data is encrypted on its own and isolated at the row level, so tenants never see each other.

Integrations

Every merchant's stack is different, so the agent works through the systems a merchant already runs rather than asking them to move.

Messaging

WhatsApp, Instagram, Messenger, TikTok, web chat

Commerce

Shopify, Salla, WooCommerce, Odoo, YouCan, Zoho

Payments

Stripe, Paymob, Fawry

Shipping

Bosta, Aramex

My role

I co-founded Zaylon and built its AI system and backend: the conversational architecture, the LangGraph agent that runs it, the tri-lingual NLP, and the FastAPI application with its database layer and the multi-tenant security around it.

• Conversational AI architecture and the LangGraph agent
• FastAPI application, services, and API surface
• Database schema, migrations, and multi-tenant isolation
• Hybrid product search: keyword and vector retrieval with rank fusion
• Tri-lingual NLP for English, Egyptian Arabic, and Franco-Arabic
• Proactive engine that spots sales opportunities in a conversation
• Auth, permissions, plan entitlements, and tenant security
• Deployment, CI, and observability for the backend

Where it stands

Integrates 6 commerce platforms and 5 messaging channels, handling product discovery, support, and checkout across three dialects. Launching commercially in 2026.