Zotlyx — Multi-Channel AI Support & Unified Inbox

Overview

Zotlyx (https://zotlyx.com) is an AI-powered customer support and unified inbox platform built for e-commerce merchants and businesses. It connects multiple communication channels—including Facebook Page comments and Messenger, Instagram DMs, WhatsApp Business, Telegram, and a website chat widget—into one shared inbox where human agents and AI assistants can collaborate.

At Yupsis, I architected the core system using a Two-Zone architecture:

1. An edge messaging zone on Cloudflare Workers to handle incoming webhooks, message grouping, and fast AI responses.
2. A relational database zone on Supabase/PostgreSQL to store store catalogs, orders, team permissions, and conversation history securely.

Architecture: Why Two Zones?

When dealing with social media webhooks, platforms like Meta and WhatsApp send massive spikes of events (for example, during holiday promotions). Processing everything synchronously inside a standard server could quickly overload the database or time out.

• Zone 1 (Edge Ingestion on Cloudflare): Inbound webhooks hit lightweight Cloudflare Workers. The workers verify HMAC signatures from Meta/WhatsApp, acknowledge the webhook immediately, and send payloads to background queues.
• Message Debouncing with Durable Objects: Real customers rarely send their question in a single message—they send multiple short bursts (e.g. "Hi", "is this available?", "in black color?"). A stateful Durable Object holds these messages for 5 seconds, combines them into one coherent question, and only then queries the AI model. This keeps AI answers relevant and avoids redundant API calls.
• Semantic Product Search (RAG): When a customer asks about a product, Cloudflare Vectorize finds the most relevant items from the store's knowledge base before drafting the reply.
• Zone 2 (Database & Admin Dashboard): Next.js dashboards connect to PostgreSQL through Cloudflare Hyperdrive connection pooling. We use oRPC for type-safe API communication between the web frontend and edge backends.

Practical Engineering Challenges & Solutions

1. Grouping Rapid Message Bursts Before Calling the AI

• Problem: If a customer sends three messages within 4 seconds, calling an LLM three separate times wastes API tokens and produces disjointed, repetitive answers.
• Solution:
  • Created a DebounceSession using Cloudflare Durable Objects with a 5-second timer.
  • Every incoming message within that 5-second window appends to an active buffer.
  • When the timer finishes with no new messages, the whole conversation chunk is sent to the AI in one prompt.
  • This cut our LLM usage costs substantially and made AI replies sound natural and complete.

2. Handling Webhook Traffic Spikes Without Dropped Messages

• Problem: During flash sales, thousands of webhooks arrive at the same time. If our backend took 2 seconds to process each one, Meta's servers would time out and stop delivering messages.
• Solution:
  • Used Cloudflare Queues to decouple webhook ingestion from business processing.
  • The edge worker acknowledges the webhook with an HTTP 200 within 15 milliseconds, placing the message in a queue for background processing.
  • This prevented any dropped inquiries even during heavy promotional spikes.

3. Understanding Mixed-Language Customer Queries

• Problem: Many customer queries in our region use a mix of local language and English (colloquial phrasing, mixed vocabulary), which standard English-only embedding models misunderstand.
• Solution:
  • Used the multilingual bge-m3 embedding model on Cloudflare Workers AI.
  • Indexed merchant FAQs, return policies, and product details into Cloudflare Vectorize.
  • The system reliably retrieves the correct context even when customer messages are informally phrased.

4. Role-Based Permissions for Growing Teams

• Problem: Merchants have different team members (support agents, store managers, billing reps) who need different access levels without complicated configuration.
• Solution:
  • Built a straightforward permission system supporting scoped permissions (like inbox:read, inbox:reply, order:manage, or * for owners).
  • Checked permissions at the middleware layer on every RPC request to ensure strict security across all team roles.

Tech Stack

Summary of Impact

• Unified Inbox: Consolidated 6 different channels into a single dashboard for support teams.
• Smart Message Grouping: Prevented fragmented AI answers by waiting for the customer to finish typing their thought.
• Reliable Ingestion: Handled high-volume webhook bursts smoothly without dropping customer messages.