Vehify was an AI sales assistant built to help businesses respond to customers through WhatsApp, Facebook, and Instagram. Instead of requiring a person to answer every message, it could recommend products, remember customer details, and book appointments inside the conversation.
I built most of the Node.js backend and React admin panel with support from the team. Vehify reached a working first version, but the company paused it before launch because of funding constraints. The source is private and the application is no longer online, so this article describes only how the system worked at a high level.
How it worked
- A business configured its workspace. It connected a supported Meta account, added company information and response guidance, uploaded its product catalog, and defined available appointment slots.
- A customer sent a message. Meta delivered the WhatsApp, Facebook, or Instagram message to the backend through a webhook.
- The backend assembled context. It identified the business and customer, then loaded relevant company instructions, catalog data, availability, conversation history, and remembered customer details.
- The AI selected the next action. OpenAI tool calling allowed it to respond normally, search the product catalog, check or book an appointment, or retrieve customer information.
- The response returned through the same channel. Product matches could include names, prices, details, and images. Booking flows collected missing contact details before confirming an available slot.
- The admin panel kept the business informed. Staff could review conversations, customer profiles, appointments, and sentiment signals, and decide when human follow-up was useful.
The engineering challenges
The central challenge was controlling AI tool use. The model needed enough context to make a useful decision without confusing a normal question with a product search or booking request. Tool inputs also needed server-side validation before they could query business data or create an appointment.
Meta integration added another boundary. WhatsApp, Facebook, and Instagram used different message shapes and account details, but the rest of the application needed one consistent conversation model. Normalizing those webhook events kept channel-specific behavior away from the shared AI and business logic.
Customer memory also had to remain practical. The system preserved useful details such as a person’s name, contact information, preferences, and earlier conversations so the assistant could continue naturally instead of starting from zero each time.
Technology
The system used Node.js, Express, React, OpenAI tool calling, and the Meta Business API. Its main domains were business onboarding, connected social accounts, catalogs, customers, conversations, appointments, and the administration dashboard.
Vehify gave me practical experience connecting AI decisions to real application actions rather than using a model only for text generation. It also reinforced the importance of validating tool calls, isolating third-party integrations, and keeping a human-visible record of automated conversations.
Vehify was a private company project. This article describes my contribution and the system only at a high level; no private source code or customer data is included.