Ozgur Karahan

AI Solution Engineer at Microsoft

Secure agentic platforms, drawn from identity to action.

I design the architecture that connects AI agents, cloud infrastructure, APIs, and the real user. The goal is simple: move enterprise AI from prototype to governed production.

See selected work Open the Agentic Platform vision
Architecture field note / identity path Live pattern
Per-user identity propagation through an agent platform A user context passes through an agent platform and policy gateway to enterprise systems while preserving authorization and traceability. User Entra ID delegated token CONTROL PLANE Agent platform Microsoft Foundry MCP tools APIM policy gateway CRM Salesforce ITSM ServiceNow APIs custom systems NO SERVICE ACCOUNT user context stays attached AUTHORIZED + TRACEABLE every action maps to a person
No service accounts Per-user authorization Traceable actions

Current focus

Selected systems, not just repositories.

Each project tests a platform-level question: how agents inherit trust, how teams ship governed software, and how knowledge compounds across tools and sessions.

Secure agent access

Agents that act as the real user.

A production pattern for carrying delegated identity from the browser through Microsoft Foundry and Azure API Management into Salesforce, ServiceNow, and custom APIs.

Read the architecture story View Entra implementation

Microsoft Foundry · Entra ID · APIM · MCP · OAuth 2.0

Challenge
Enterprise agents must not hide behind shared credentials.
Move
Propagate the user's delegated identity across every boundary.
Proof
Working patterns for Entra, Salesforce, ServiceNow, and custom APIs.

Governed software delivery

From workplace request to deployed software.

A presenter-led Agentic SDLC demo that connects intake, planning, implementation, testing, review, pull request, and deployment with enforce-first CI gates.

Explore the project Open the lifecycle deck

GitHub Copilot · GitHub Actions · Multi-agent workflow · JavaScript

Challenge
Agentic coding needs governance without losing delivery speed.
Move
Make policy and verification part of the workflow, not an afterthought.
Proof
A complete, repeatable path from request intake to deployment.

Durable agent knowledge

Memory that compounds across sessions.

A Karpathy-style LLM wiki that gives coding agents durable project and domain context across GitHub Copilot, Claude Code, Codex, and Cursor.

Explore AI Agent Memory

Python · Knowledge graph · GitHub Copilot · Claude Code · Codex · Cursor

Challenge
Agents repeatedly relearn the same project context.
Move
Turn lessons, decisions, and terminology into a linked local wiki.
Proof
One memory pattern shared across multiple coding-agent runtimes.

The Agentic Platform: our vision.

An interactive deck on the shift from AI assistance to an operating platform for agents. It follows the full system from how developers build agents to how enterprises deploy, operate, and govern them at scale.

Open the interactive deck View source on GitHub
  1. Build skills · tools · memory
  2. Deploy runtime · identity · policy
  3. Operate traces · evaluation · reliability
  4. Govern access · cost · accountability

Working across the whole system.

My work sits between strategy and shipping. I translate enterprise constraints into platform decisions, then carry those decisions into code, infrastructure, and demonstrable systems.

Agent systems
Microsoft Foundry, MCP, Claude SDK, A2A, memory, evaluation, and multi-agent workflows
Trust and integration
Entra ID, OAuth 2.0, API Management, Salesforce, ServiceNow, MuleSoft, and custom APIs
Cloud delivery
Python, React, FastAPI, Bicep, Container Apps, GitHub Actions, KQL, and observability

Complete project index.

Selected work above carries the story. This archive keeps the full technical record, including public repositories, contributions, learning work, and safely redacted private projects.

Browse all 15 projects 2 private repositories are explicitly flagged

Agent platforms and delivery

  • Public project

    AI Agent Memory

    ai-agent-memory

    Persistent memory for AI coding agents using a linked, local LLM wiki pattern.

    Python · GitHub Copilot · Claude Code · Codex · Cursor

  • Public project

    Agentic SDLC Demo

    agentic-sdlc-demo

    A complete presenter-led software lifecycle driven by coding agents and enforce-first CI gates.

    JavaScript · GitHub Copilot · GitHub Actions · Multi-agent

  • Internal template Private repository

    AI Agent Project Template

    An internal scaffold that standardizes project structure and agent-configuration conventions.

    Agent configuration · Project conventions

  • Learning project

    Anthropic Academy

    anthropic-academy

    Worked examples from the Anthropic Academy curriculum.

    Python · Claude API

Identity and integration

  • Public project

    ServiceNow Meta-tool

    snow-meta-tool

    Three generic MCP tools for discovery, query, and write operations with per-user identity.

    Microsoft Foundry · ServiceNow · MCP · APIM · Bicep

  • Public project

    Salesforce Identity Propagation

    salesforce-meta-tool-identity-propagation

    Cross-cloud identity propagation from Azure to Salesforce through an MCP meta-tool.

    Microsoft Foundry · Salesforce · MCP · OAuth 2.0

  • Public project

    Entra Identity Propagation

    propagate-id-entra

    End-to-end identity propagation from browser to agents, policy gateway, and backend APIs.

    Microsoft Foundry · Entra ID · APIM · MCP · Bicep

  • Public project

    Simple Order Agent

    simple-order-agent

    An order-processing agent that connects Claude SDK, MuleSoft MCP, and the A2A protocol.

    Python · Claude SDK · MuleSoft · MCP · A2A

  • Public project

    Azure APIM MCP Server

    azure-apim-mcp-server

    A streamable HTTP MCP server for managing Azure API Management through natural language.

    Python · Azure API Management · MCP

Applied AI and operations

  • Public project

    Azure SRE Agent for Synapse

    azure-sre-agent-synapse-poc

    Reproducible infrastructure, KQL runbooks, and smoke tests for Synapse observability scenarios.

    Bicep · KQL · Azure SRE Agent · Synapse

  • Public project

    AI Avatar Presenter

    ai-avatar-presenter

    A multilingual avatar presentation assistant with slide upload, speech, and interactive Q&A.

    React · FastAPI · Azure AI Speech · Azure OpenAI · WebRTC · Bicep

  • Enterprise delivery Private repository

    Enterprise Document RAG

    Document intelligence and retrieval-augmented generation delivered for an enterprise client.

    Python · Azure AI · RAG

Contributions and references

  • Extended fork

    Get Started with AI Agents

    get-started-with-ai-agents

    Azure Samples fork extended with Foundry endpoint, A2A documentation, Bicep, and setup guidance.

    Python · Microsoft Foundry · Bicep

  • Reference fork

    Voice Live Sales Coach

    voicelive-api-salescoach

    AI-powered voice training for sales professionals, based on an Azure Samples reference.

    Python · Azure AI Speech · Voice Live API · React

  • Reference fork

    Karpathy Autoresearch

    karpathy-autoresearch

    An autonomous research agent that runs single-GPU LLM training experiments overnight.

    Python · PyTorch · LLM training

Enterprise AI architecture · agent platforms · identity

Have a platform problem worth drawing out?

Start a conversation Explore GitHub