Sellavie.ai: the sales agent is only half the system

The tempting way to describe Sellavie is simple: an AI sales agent for Instagram, Facebook, and WhatsApp.

That is true, but it is also the least interesting part of the build.

The harder part was making the agent belong inside a business system. A customer message is not just text. It might belong to one business, one product catalog, one pending order, one payment path, one human handoff, and one set of rules the AI is not allowed to casually cross.

The actual problem

Most AI demos stop after a good response.

Sellavie has to keep going after the response. It needs to know which business is speaking, what products are available, what the customer is trying to do, and when the system should stop acting confident and bring a human into the loop.

That changes the engineering shape. The LLM becomes one component inside a larger workflow, not the center of the whole product.

What I built around the AI

I worked on Sellavie as a full product system: backend services, web dashboard, mobile surfaces, business-scoped access, product catalog flows, training controls, subscription guardrails, analytics, and operational tooling.

The backend owns the sensitive decisions. The dashboard gives business users control without exposing secrets or unsafe configuration. The mobile app keeps the operational work reachable when the owner is not sitting at a laptop.

The important design choice was separation:

• customer-visible replies are one layer
• business workflow execution is another layer
• provider events and payment-adjacent state changes are handled server-side
• team access, handoff, and audit trails sit outside the AI response itself

That separation makes the product less flashy on paper, but much more useful. An AI that can confidently answer a customer question and then accidentally re-open a closed order is not a product. It is a liability.