The problem

Most storefront chatbots can only answer questions about the store. They cannot tell a customer where their order is, because they are not connected to anything: the moment a conversation needs a real fact or a real action, it turns into a support ticket and the customer waits. Debo Web wanted the opposite — an assistant that could actually do the job, on the storefront, at the moment the customer asks.

My role

Software Architect and AI Automation Engineer. I architected the system and shipped it: the agent design, the LLM orchestration, and the automation architecture that connects it to live commerce operations.

Architecture

• An agentic assistant rather than a scripted bot: it handles order status, payment assistance and customer support directly on the storefront, taking the steps a support agent would take instead of routing to one.
• LLM workflows orchestrated across the OpenAI, Anthropic Claude and Google Gemini APIs — no single-vendor dependency, and each step routed to the model that suits it.
• The automation layer is what makes it real: webhooks, queues and explicit API contracts connecting the assistant to orders, payments and notifications. Queues matter more than they look here — a conversation should never block on a slow downstream system, and a failed action has to be retryable rather than lost in a chat log.
• n8n and Zapier carry the sales and operations automations, which keeps those workflows editable by the business instead of locked inside application code that needs a developer for every change.
• Automated lead-capture and follow-up sequences on the sales side, so a conversation that is not a support request still ends somewhere useful.

Technologies

OpenAI API, Anthropic Claude API, Google Gemini API, n8n, Zapier, webhooks and queue-based integration.

Outcome

The assistant went to production and handles order status, payment assistance and customer support on the storefront. The design lesson I carry forward is that the model is the least interesting part of an AI product. What decides whether the feature is useful is the integration surface behind it: clean contracts, idempotent actions, retries, and a firm boundary between what the agent is allowed to decide and what it is allowed to execute. Three-month engagement, January to March 2025.