Enterprise AI Platform

The problem

Every business wants “an AI assistant for our company”, but the hard parts — grounding, security boundaries, audit, voice and approvals — are the same every time. Rebuilding them for each project is slow and makes each one a new risk.

What I built

• A fixed shared core (AI, messaging, web, agents) with a strict dependency rule: business modules reference only the core, industry editions add only their differences, and only a host composes them — enforced by build-breaking tests.
• Six web hosts on the same core: a commercial website; hotel, real-estate, healthcare and commerce demos; and an internal finance application.
• RAG over Azure AI Search with relevance gates and numbered citations. Untrusted text enters a prompt only through one fencing routine, and tools and indexes are chosen on the server, never by the browser.
• Live voice on Azure OpenAI Realtime over WebRTC, sharing one conversation with text chat; short-lived client secrets mean the main key never reaches the browser.
• Two orchestration engines behind one setting: a hand-written pipeline and Microsoft Agent Framework.
• Lead capture, conversation wrap-up and notifications; protected administration with audit trails and human approval steps.

Architecture

1. • Hosts

     website · hotel · real estate · healthcare · commerce · finance

2. • Industry editions

     only the differences: persona, knowledge, rules

   • Business modules

     accounting · marketing · commerce

3. • Shared core

     AI · RAG · realtime voice · agents · messaging · web

4. • Azure

     OpenAI · AI Search · SQL · App Service · Key Vault

Key decisions

• One rule, one place

  A rule restated in each channel’s own words is two rules that will drift apart. Safety rules are one runtime value that every prompt — text, retrieval and live voice — emits word for word.

• Boundaries are code, not prompts

  Which index a public assistant may read is a positive allow-list checked at startup; a misconfigured deployment refuses to start instead of exposing private data.

• Risky features ship switched off

  Anything that spends money, accepts files from strangers or reaches a model is off by default and enabled per deployment, with the default pinned by a test.