Yann AbadieHire me on Upwork

AI agents in real systems, safely.

The agent does the work, a person approves what matters (proof: Meta-YGN README, a risky command creates a checkpoint and then asks), and every action is traceable (proof: Confidant README, a receipt with a SHA-256 for every outbound call). I build that layer for agents and MCP servers on Azure and Microsoft 365.

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.

Illustration: replays a typical run

  1. 14:02:07agentread_file src/billing.pyallowed read scope
  2. 14:02:08agentsql.query SELECT … LIMIT 50allowed read-only scope
  3. 14:02:10agentshell rm -rf ./datablocked risk 0.94, checkpoint created
  4. 14:02:11agentmail.send to=finance@example.comheld waiting for a person
  5. --:--:--youmail.send to=finance@example.compending your decision
  6. --:--:--logsha256 of this log, written after your decision
Approve and Reject need JavaScript. The run above is complete.

I build this layer between your agents and your systems. Each part is proven in a public repository:

  • Risk scoring and blocking before a command runs: Meta-YGN.
  • Your own test suite decides whether an agent’s fix passes: SwarmProof.
  • A second model checks each tool result: nexus-swarm-os.

Selected work

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.

  1. Meta-YGN

    Status: Released

    A Rust daemon for Claude Code that blocks destructive commands before they run (proof: Meta-YGN README, line 7), after taking a checkpoint.

    Rust, tree-sitter, SQLite FTS5, TypeScript, MCP. Last commit 19 November 2026.

  2. SwarmProof

    Status: Alpha

    Claude and Codex propose a fix, your own test suite gives the verdict (proof: SwarmProof README, line 9-12), and a hashed Evidence Pack records it.

    Python 3.11 standard library, Claude CLI, Codex CLI, GitHub Actions. Last commit 5 November 2026.

  3. Confidant

    Status: Hackathon

    A personal assistant where the cloud model never sees a private fact (proof: Confidant README, line 7-8). It writes placeholders, and the device fills them in.

    Python, SQLite FTS5, Nemotron on Nebius Token Factory, MCP, Gradio. Last commit 29 October 2026.

  4. nexus-swarm-os

    Status: Working

    Multi-model orchestration in which the opposite model checks each tool result (proof: nexus-swarm-os README, line 19) before the pipeline moves on.

    Python 3.11+, 7 provider SDK drivers, Docker, FastAPI, MCP. Last commit 19 November 2026.

  5. Factur-X France 2027

    Status: Released

    A zero-dependency Python validator for French e-invoices, shipped as a CLI, a library and a GitHub Action.

    Python standard library, GitHub Actions, PDF/A-3, EN 16931. Last commit 17 November 2026.

Also on GitHub

  • mnemo ties a prepared decision to the facts behind it and refuses a stale export.
  • mcp-session-check checks MCP HTTP discovery and the session lifecycle without calling any tool.
  • oida-code gives diagnostic evidence for AI-written Python diffs.
  • metascaffold is a metacognition MCP server for Claude Code.

Ways to engage

All 3 run on Upwork: hourly at $90/h, or a fixed price agreed on Upwork before work starts.

Architecture review

Paid consultation, fixed scope

Your team wants an agent with access to code, databases or mail, and nobody can say yet what it could break.

What ships

  • A map of every tool the agent can call and the credential behind it
  • A decision for each action: allow, block, or hold for a person
  • The changes that remove the most risk, in order
  • A written report your team keeps

How it starts: Book the consultation on my Upwork profile.

Agent or MCP build

Milestones on Upwork

You need an agent or an MCP server working against your own systems, with approvals and logs from the first release.

What ships

  • An agent or MCP server with least-privilege, revocable credentials
  • An approval step on every write action
  • An action log with hashes you can re-check
  • Tests in your CI and a runbook for the people who operate it

How it starts: Invite me to a job on Upwork; we agree the milestones there.

Microsoft 365 and Entra security review

Fixed scope on Upwork

An AI assistant or connector needs Microsoft 365 data, and you want its access limited to what it needs.

What ships

  • Conditional Access policies checked for the identities your connectors use
  • App registrations, permissions and consent compared with what each app needs
  • Graph mail access scoped with Exchange Online RBAC to the mailboxes an app needs
  • Findings ranked by impact, each with its fix

How it starts: Start with the consultation, or invite me to a fixed-scope job on Upwork.

How I work

A first engagement runs in this order.

  1. Least privilege

    We start with access. The agent gets the narrowest credential that does the job: read-only first, one mailbox, one schema.

  2. 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.

  3. Your tests decide

    An agent saying “fixed” proves nothing. Your test suite, restored before each run, gives the verdict.

  4. 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.