AI Solution Engineer at Microsoft
Useful agents need more than a model.
I build and share enterprise agent patterns where business workflows, user identity, and human decisions meet.
Explore the work and ideas · Connect on LinkedIn
| Question | What I explore |
|---|---|
| What can it change? | A concrete business action, not just a conversational answer. |
| Whose authority does it use? | Caller identity and authorization across system boundaries. |
| When should a person decide? | An explicit mandate, escalation condition, and human handoff. |
These are design questions, not a claim that the separate projects below form one production system.
A real recorded phone call reaches a Foundry voice agent through Voice Live and Azure Communication Services. The agent checks a booking, quotes a change before acting, and routes a discount outside its mandate to a human in Teams through Power Platform.
Watch the demonstration — French, English subtitles, 6m19s · Source · Presentation and additional resources
Scope: mock CRM, ERP, and knowledge-base systems, not live Dynamics or SAP integrations. This is a demonstration, not a production service or a public phone line to try.
Separate implementations explore caller identity through Entra ID, Foundry, APIM, and downstream systems. Propagating an identity does not, by itself, establish appropriate authorization.
Entra ID proof of concept · Salesforce sample · ServiceNow sample · Evaluation boundaries
The Entra project is a proof of concept; Salesforce and ServiceNow are sample implementations, not audited production guarantees.
An experimental community sample combining the OpenAI Agents SDK, a terminal, persistent workspace, browser verification, and scoped Azure access.
Project overview · Source and walkthrough · Architecture · Limitations
Not an official Microsoft product or reference implementation, not production-ready, and no SLA or bundled infrastructure provisioning.
- The Agentic Platform — an interactive presentation connecting how agents are built, deployed, operated, and governed.
- Identity propagation: the architecture story — the reasoning behind the Salesforce meta-tool and identity flow.
- AI Agent Memory — a local Markdown wiki for coding-agent knowledge across projects and sessions, not a customer-data memory solution.
- Agentic SDLC Demo — a presenter-led delivery demo and harness. Its README documents current setup and status; this is not a claim of a complete, verified workflow. Presentation · Companion repository.
Other projects, presentations, and private work
| Project | Focus |
|---|---|
| Azure APIM MCP Server | Exposes a business orders REST API as MCP through APIM's native gateway; Python, FastAPI, PostgreSQL, and Bicep. |
| Simple Order Agent | Order-processing example using Claude SDK, MuleSoft MCP, and A2A. |
| AI Avatar Presenter | Avatar-led slide presentation and interactive Q&A. |
| Azure SRE Agent for Synapse | Proof-of-concept scaffold for investigating Synapse observability scenarios. |
| Anthropic Academy | Worked learning examples. |
| The Agentic Platform | A presentation exploring the shift from AI assistance to agent platforms. Source. |
| Enterprise Document RAG | Private work. |
| AI Agent Project Template | Private work. |
Modified and reference forks, with upstream attribution
These projects originate with the upstream authors linked below.
| Fork | Upstream | Purpose |
|---|---|---|
| Get Started with AI Agents | Azure-Samples | Modified fork with endpoint/A2A documentation and infrastructure customizations. |
| Get Started with AI Chat | Azure-Samples | Chat application reference. |
| Voice Live API Sales Coach | Azure-Samples | Voice-training reference. |
| L300 AI Apps and Agents Workshop | Microsoft | Workshop reference. |
| Interactive Prompt Engineering Tutorial | Anthropic | Learning reference. |
| Karpathy Autoresearch | karpathy | Autonomous research reference. |
AI agents and apps · Identity and security · Learning resources
For architecture discussions, speaking opportunities, and professional connections, find me on LinkedIn. To reproduce a sample, report a limitation, or propose an integration, start with its repository and contribution guidance.