I design and build AI-powered applications on Azure, with deep expertise in AI agents and MCP (Model Context Protocol) integrations, and a strong background in enterprise architecture, integration strategy, and API management. My work focuses on making AI practical and production-ready for enterprise environments.

A complete voice-agent customer-service demo, built in 3 days: a customer calls a real phone number on a Saturday, an Azure AI Foundry hosted agent answers in French, verifies identity, checks real capacity, quotes before acting, executes the change in live (mock) CRM/ERP systems, sells an option — and escalates the discount decision it has no mandate for to a human in Teams. One agent, two experiences: transform experiences, not systems.

🎥 Watch the real call (French, EN subtitles): ozgurkarahan.github.io/horizon-famille · 📽️ Interactive deck · PDF

📄 The story on LinkedIn: my post about "the other ending"

💻 Code (MIT): horizon-famille — agent, mock ERP/CRM/KB, ACS⇄Voice Live bridge, Power Platform approval, deck

Enterprise AI agents must act as the real user — no service accounts in the data path, every action traceable and authorized per-user. This is a series of production-grade implementations of that pattern on Azure (Entra ID, APIM, AI Foundry) across ServiceNow, Salesforce, and custom APIs.

📄 Read the story: From Theory to Production: Salesforce Meta-Tool & Identity Propagation

🧩 The series: Entra ID Identity Propagation · ServiceNow MCP Meta Tool · Salesforce MCP Meta Tool

Projects I built from scratch.

Voice Agent	Horizon Famille — « L'autre fin » ⭐ Full voice-agent customer-service demo built in 3 days — real phone call, French voice agent on Foundry, live mock CRM/ERP, human-in-the-loop approval in Teams (🎥 demo & slides)
Agent Memory	AI Agent Memory ⭐ Persistent memory for AI coding agents — Karpathy LLM Wiki pattern for GitHub Copilot, Claude Code, Codex, and Cursor
Agentic SDLC	Agentic SDLC Demo Ready-to-run, presenter-led demo of the full SDLC driven by AI coding agents — intake → plan → implement → test → review → PR → deploy, with enforce-first CI gates (🎞️ slide deck · live validation target)
Agent + MCP	ServiceNow MCP Meta Tool ServiceNow MCP meta-tool — 3 generic tools (discover, query, write) with per-user identity propagation via Azure APIM JWT Bearer OBO flow
Agent + MCP	Salesforce MCP Meta Tool Meta-tool MCP server for Salesforce with cross-cloud identity propagation — AI agents manage Salesforce resources using the caller's Azure identity end-to-end
Agent + MCP	Entra ID Identity Propagation End-to-end Entra ID identity propagation from browser through AI Foundry agents and APIM to backend APIs — no service accounts in the data path
Agent	Simple Order Agent Order processing agent using Claude SDK, MuleSoft MCP, and A2A protocol
MCP	Azure APIM MCP Server MCP server for Azure API Management — manage APIs through natural language
SRE Agent	Azure SRE Agent for Synapse Reproducible IaC, KQL runbook, and smoke-test scaffold for validating Azure SRE Agent against common Synapse observability pain points
Template	AI Agent Project Template Internal scaffold standardizing AI-agent project structure and agent-config conventions across repos

AI App	AI Avatar Presenter AI-powered avatar presentation assistant — upload PowerPoint slides and have an AI avatar present them with multilingual TTS, interactive Q&A, and Teams integration
AI App	Enterprise Document RAG Document-intelligence and retrieval-augmented generation solution delivered for an enterprise client

Identity propagation patterns — see full project descriptions in AI Agents & MCP above.

Learn

Anthropic Academy Worked examples from the Anthropic Academy curriculum

Vision Deck

The Agentic Platform — our vision Interactive vision deck on the shift from AI assistance to agentic platforms — how developers build, deploy, and operate agents at scale (source)

Fork+

Get Started with AI Agents  ·  forked from Azure-Samples My additions: AI Foundry endpoint & A2A documentation (1 400+ lines), Bicep infra customizations, extended README

Unmodified forks kept for learning and workshops.

	Project	Upstream	Purpose
	Get Started with AI Chat	Azure-Samples	Azure AI Foundry chat application reference
	Voice Live API Sales Coach	Azure-Samples	AI-powered voice training for sales with Azure Voice Live API
	L300 AI Apps and Agents Workshop	Microsoft	L300-level AI workshop material
	Interactive Prompt Engineering Tutorial	Anthropic	Interactive prompt engineering exercises
	Karpathy Autoresearch	karpathy	Autonomous AI agent running LLM training experiments

Enterprise identity propagation patterns for multi-cloud AI agent architectures — making every AI action traceable to the real user across Azure, Salesforce, and ServiceNow.

All repositories follow a standard convention for working with AI code agents:

• AGENT.md — public project instructions, architecture, and conventions committed to each repo
• .claude/ — private per-project configuration, lessons learned, and session context (gitignored)
• Global workflow rules — verification standards, subagent strategy, and self-improvement loop loaded across all projects

This separation keeps project knowledge accessible to collaborators while keeping iterative agent context private.

Currently exploring enterprise identity propagation patterns for multi-cloud AI agent architectures.