Highlighted phrases link to the public README line that proves them.
Role
Senior AI & Cloud Security Architect
Based
France
Hours
It is in Paris. I work evenings and weekends, Paris time.
Rate
$90/h, contracted and paid on Upwork
Works with
AI agents, MCP servers, Azure, Microsoft 365, Entra ID
The approval gate
An agent proposes tool calls. The gate allows reads inside the agent’s scope, blocks a destructive command after taking a checkpoint, and holds any call that sends, pays or deletes until a person decides. The log ends with a hash of every line above it.
5 public repositories. Highlighted claims link to the README line that makes them, every number comes from those READMEs, and each status says how far the project got.
We start with access. The agent gets the narrowest credential that does the job: read-only first, one mailbox, one schema.
Human approval on writes
Then the gate. Anything that sends, deletes, pays or deploys waits for a named person, and the agent cannot approve itself.
Your tests decide
An agent saying “fixed” proves nothing. Your test suite, restored before each run, gives the verdict.
Evidence you can re-check
Every run leaves a log with SHA-256 hashes, so you can verify what happened without taking my word for it.
About
About 10 years in IT. Client work stays confidential. Everything on this site is public and checkable.
Career
Now
Solutions Architect in an industrial manufacturing group: integration projects and the IT department’s AI program.
Before
Head of Cloud Architecture & Security.
Earlier
Microsoft Cloud & Security Architect: Azure, Microsoft 365, Entra ID, Intune, Defender, Zero Trust.
Recent work, without names
Built an MCP server that gives an AI assistant read-only access for code analysis.
Wrote an AI-agent governance framework aligned with OWASP guidance on agentic AI.
Designed an offline, on-premises AI server for sensitive data.
Secured Microsoft 365 access for AI connectors: Conditional Access, and Exchange Online RBAC that limits Graph mail access to one mailbox.
Languages
French (native), English (fluent)
Hours
Evenings and weekends, Paris time
Also work with
Terraform, Sentinel and KQL, Kubernetes and Rancher, Dynamics 365, MIM, LangGraph
Questions
Why Upwork only?
One place for the contract, milestones, invoices and payment. Upwork holds fixed-price milestone funds in escrow while you review the work, and both of us get its dispute process. Send me a message there first.
Do you work with local or on-prem models?
Yes. I designed an offline, on-premises AI server for sensitive data. nexus-swarm-os has an Ollama driver for local models, and Meta-YGN runs on your machine with no network calls.
How do you handle access to our data?
I ask for the narrowest access that does the job, in accounts you own and can revoke. Write actions go through an approval step, logs stay in your systems, and client work stays confidential.
Can you start small?
Yes. The paid consultation on Upwork is a fixed-scope first step: one system, one agent, one written result. If it is useful, we plan milestones from there.
What about time zones?
I work evenings and weekends, Paris time. For most of the year, 18:00 to 23:00 Paris time is the working afternoon on the US East Coast and the morning on the West Coast.