Multi-Tenant AI Platform

the problem

A clinic or a salon wants an assistant that answers customers on the web, Telegram and WhatsApp from its own data. Each business has to stay completely separate from the others.

the approach

• Split the system in two. A Laravel application is the system of record for tenants, branches, catalogues, bookings, complaints, FAQs and broadcasts. A separate Python service runs the reasoning agent and keeps no state of its own.
• The agent reaches business data only by calling back into Laravel through tool endpoints signed with HMAC, so every answer passes the same tenant and role checks as the web application.
• Retrieval runs on a vector database. Before choosing one, I built a small tool to compare two candidates on the same documents.
• Guardrails keep the assistant on topic and refuse what falls outside a business, and answers stream to the customer as they are generated.

the outcome

• The company's first in-house product, designed and built by one engineer.
• One assistant per business across web, Telegram and WhatsApp, with tenants isolated from each other.
• Automated tests and evaluation reports on both the PHP and the Python side.