Daniel Ferreira / Home

Selected work / SKF

Agentic SOC.

Human judgement. Amplified.

At SKF, I built an identity-first Agentic SOC platform and internal Cyber MCP, bringing specialised agents and documented procedures to security engineering and operations. Less repetitive work. More context for the analyst. Consequential decisions stay with people.

Follow the architecture

Architecture overview · illustration only. This page is not connected to operational systems.

Delivery, with control.

Reported impact across security engineering and operational delivery at SKF.

Security engineering & operational delivery increase
~5x
AI-enabled SOC operations
24/7
Password-less, governed access
Identity-first

Two gateways. Different jobs.

The model gateway routes reasoning requests; the Function gateway controls tool execution. Model output grants no system access. Evidence returns to people for review.

Reasoning → governed execution → analyst review.

  1. Human direction → model reasoning

    1. Human

      Sets intent and boundaries.

    2. OpenCode + Codex

      Harnesses coordinate agents and procedures.

    3. Model API gateway / OmniRoute

      Model routing and token optimisation.

    4. Model intelligence

      OpenAI / Anthropic reasoning, including Azure Foundry models.

  2. Governed tools → accountable operations

    1. Governed MCP / tools

      Harnesses request scoped operations.

    2. Entra-gated Function gateway

      Audit first; execute only if the record succeeds.

    3. Systems

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

    4. Evidence / human-approved actions

      Reports and recommendations; people approve consequential actions.

Six capabilities. One operating model.

From alert to analyst.

Illustrative workflow · from evidence to a human decision.

  1. Receive

    A Sentinel alert supplies the starting entities and investigation question.

  2. Investigate

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

  3. Enrich

    Device posture, identity context and threat intelligence inform the assessment.

  4. Document

    Evidence, uncertainty and suggested next steps accompany the incident.

  5. Review

    An analyst checks the findings and approves the response; no automatic containment is implied.

Control is part of the design.

What comes next.

Pending roadmap

  • Network integrations

    Zscaler ZIA / ZPA and Palo Alto connector integrations for traffic, policy and firewall analysis remain pending.

  • Active containment

    Blocking indicators, acquiring or removing files, email quarantine and device isolation remain pending, with human approval required.

The original blueprint also marks OT vulnerability coverage, CyberArk identities and OT/AWS/legacy estate expansion as pending.

Better tools. Accountable teams.

Discuss the work