Daniel Ferreira / Home

Selected work / SKF

Agentic SOC.

Security operations, with people in control.

At SKF, I built an identity-first platform and internal Cyber MCP to connect specialised agents, documented procedures and security tools. The aim: give analysts useful evidence and reduce repetitive work, while keeping consequential decisions with people.

Explore the system

Illustrated architecture · not connected to operational systems

The work in context.

Owner-reported impact and operating model at SKF.

Increase in security engineering and operational delivery
~5x
AI-enabled security operations
24/7
Password-less access to governed tools
Identity-first

Reasoning and execution have separate routes.

The model gateway routes requests for reasoning. A separate Function gateway governs tool execution. Evidence comes back to analysts, who review the findings and any consequential action.

A model can recommend an action; it cannot grant itself access.

  1. Direct and reason

    1. Human

      Sets the question and boundaries.

    2. OpenCode + Codex

      Coordinate specialised agents and documented procedures.

    3. AI API gateway / OmniRoute

      Routes models and helps optimise token use.

    4. Model intelligence

      OpenAI and Anthropic models, including through Azure Foundry.

  2. Access tools and review

    1. Governed MCP / tools

      Agents request operations within their scope.

    2. Entra-gated Function gateway

      Records the request before execution; stops if audit logging fails.

    3. Systems

      Sentinel, Defender, Entra, Azure, ServiceNow, SharePoint and Outlook.

    4. Evidence / human-approved actions

      Findings return to people; consequential actions require approval.

How the routes connect

  1. OpenCode + CodexRequests scoped toolsGoverned MCP / tools
  2. Evidence / human-approved actionsEvidence returns for reviewHuman

Where the platform is used.

An investigation, step by step.

Illustrative example · the analyst remains responsible for the decision

  1. Receive

    A Sentinel alert provides the entities and the question to investigate.

  2. Investigate

    Agents follow documented procedures and request identity-gated, audited queries.

  3. Enrich

    Device posture, identity context and threat intelligence add context.

  4. Document

    The incident record brings together evidence, uncertainty and suggested next steps.

  5. Review

    An analyst reviews the findings and approves the response. The example does not imply automatic containment.

Boundaries built into the workflow.

Interested in the work?

Discuss the work