AI operations / Workflow engineering & governance

AI Systems & Agent Operations

7 min read
  • AI Workflow Automation & Systems
  • Sales Enablement & SOPs
  • Go-to-Market & Growth Strategy

Designing reliable foundations for AI-assisted operations—with human approval, knowledge governance and security built in.

Focus
AI operations / Workflow engineering & governance
Core skills
AI Workflow Automation & Systems · Sales Enablement & SOPs

Alongside my work in marketing and growth, I design and develop AI-assisted systems for content operations, knowledge management, workflow coordination, and digital product development. My work combines Hermes Agent, large language models, Python-based tooling, cloud infrastructure, structured knowledge records, and human-approval controls.

The focus is not simply using AI to generate content. It is designing governed workflows in which AI can research, draft, coordinate, and surface decisions while respecting clear approval boundaries, security controls, and rules against unsupported claims.

Hermes Agent & AI Operations

I installed and configured Hermes Agent across a Windows 11 workstation and a separate Tencent Cloud VPS environment. I worked with agent sessions, persistent memory, custom skills, scheduled jobs, browser automation, delegated sub-agents, configuration files, service management, diagnostics, and backups.

I also reviewed model-provider configuration and API connectivity, checked the health of the agent environment, and worked through the relationship between language models, tools, operating-system services, and cloud-hosted execution. The VPS environment was configured to run Hermes as a system service, providing an always-on cloud-hosted agent environment alongside the workstation setup.

Supplied console view of the cloud-hosted Hermes Agent instance: status Running in the Singapore region, Starter Type specifications of 2 vCPUs, 2GB memory, 40GB Cloud SSD and 512GB/month transfer. The network address and instance ID are redacted.

Autonomous Content Orchestrator — Design & Foundation

I designed a multi-stage workflow for Node Digital's own content operations, spanning content planning, production, approval, publishing, performance measurement, and learning. The planned channels include Facebook, Instagram, LinkedIn, X, TikTok, the website/blog, and SEO content. The scope is limited to Node Digital's own channels, not client accounts or client operations.

The work focused on establishing a controlled workflow before expanding automation: defining stages, clarifying responsibility, validating assumptions in read-only mode, and setting explicit approval points for consequential actions. This is an orchestration design and foundation build; it should not be represented as a fully deployed, end-to-end autonomous publishing system.

Supplied session view of the orchestrator's content-engine rules: approval and change-control handling for content, creative assets, calls to action, links and publication timing, including the material-change rule that voids an earlier authorization. This documents the workflow design, not a live publishing system.

Knowledge & Memory Architecture

I designed a four-tier knowledge model:

  • Agent memory
  • Custom skills
  • Google Drive as the intended canonical knowledge store
  • Session scratch space

The design includes seven knowledge categories, provenance information for records, a machine-readable registry, default-closed approval for public use, and a Proof-Point Register. The purpose is to preserve context, distinguish source-backed facts from assumptions, and prevent metrics or claims from entering public content until their wording has been owner-verified.

Google Workspace/Google Drive was designed as the canonical store, but the connection status should be verified before describing that integration as fully operational.

Orchestrator Control Foundation & Security Boundary

I worked on a Python and SQLite coordination foundation for reliable agent workflows. The design and implementation include a versioned schema, durable transaction lifecycle, idempotency and conflict detection, compare-and-set versioning, handoff and reconciliation records, an audit outbox, restart recovery, and unit tests.

On Windows, I also developed a Writer/Coordinator security boundary using an authenticated named pipe and a separate, least-privilege service account. The coordination database is protected with Windows access-control permissions so the interactive user cannot simply bypass the approved write path. The system is designed to fail closed when the required controls are not satisfied.

This work applies software engineering and security principles to agent operations: controlling how changes are written, ensuring that operations can be audited, and making recovery and conflict handling part of the design rather than an afterthought.

Scheduled Agent Operations & Custom Skills

I configured a daily Knowledge Digest cron workflow intended to surface unresolved questions, low-confidence information, and high-risk blockers for review instead of allowing the agent to act on uncertainty.

I also developed Node Digital-specific Hermes skills covering content workflows, the knowledge layer, image generation, media production, and orchestrator planning. These skills encode brand standards, approval requirements, and client-confidentiality rules so that the same operating principles can be applied consistently across sessions.

Supplied text diagrams of the deployment: a cloud server running 24/7 with Hermes, the Orchestrator, Writer and Durable State reaching Google, AI APIs and Publishing, and a Windows PC where Hermes UI and Administration both connect to the cloud. These diagrams document the intended layout of the environments.

AI-Assisted Ebook & Content Development

I used AI-assisted research and drafting workflows for long-form marketing content and a full-length ebook project with the working title Meta Lead Ads A–Z. The workflow used a Markdown master document followed by DOCX production, kept source notes separate from reader-facing text, and included image placeholders for designer handoff.

I have also used AI to help organise marketing case studies, professional achievements, educational content, and editorial material. The workflow combines research and drafting with source separation, content structure, review, and preparation for downstream design. The ebook's publication and sales status should be stated separately; this work does not imply that it is currently on sale.

AI-Assisted Website Planning

I developed a platform-neutral Master Website Brief and platform-specific website build guidance for Node Digital. This work covered information architecture, brand positioning, page structure, portfolio case studies, SEO content, future ebook functionality, authentic asset use, and requirements for honest handling of incomplete features.

Lovable was explored as a proposed AI website-building workflow, and Squarespace Blueprint AI was also evaluated. The specifications include gap analysis where a platform may not meet the full requirements. These activities demonstrate AI-assisted product planning and implementation preparation, not a claim that the finished website has been deployed through either platform.

Operating Principles

The work is guided by a small set of controls:

  • Human approval for consequential actions: the agent proposes; the owner approves.
  • No invented evidence: do not fabricate facts, testimonials, clients, metrics, or results.
  • Client confidentiality: client names and identifying details are excluded from public content and shared knowledge records.
  • Validation before implementation: prefer read-only inspection before changes, followed by read-back verification.
  • Secure credential handling: secrets should not be pasted into chat; accounts and purchases remain owner-controlled.
  • Traceability and recovery: retain provenance and audit information, handle conflicts, and design for restart recovery.

Language models recorded in Hermes session history

  • `anthropic/claude-fable-5.1` — recorded as the primary model through the Nous inference API.
  • `openai/gpt-5.6-luna` — recorded as used for Orchestrator and ebook sessions through the Nous inference API.

These are the model identifiers recorded in the working environment; they are presented as recorded identifiers rather than generalised product names.

Auxiliary models and providers configured in Hermes

  • `whisper-1` — configured for speech-to-text; actual usage is not verified in the available records.
  • `veo3.1` — configured for video generation; actual usage is not verified in the available records.
  • Image-generation provider — configured, but the specific model identifier is not recorded.

Other AI tools

  • ChatGPT / OpenAI — used in broader AI-assisted research, writing, planning, and ebook/Canva workflows; the exact model variants used in those separate sessions are not established in the project file.
  • Lovable — explored for AI-assisted website development and build planning.
  • Squarespace Blueprint AI — evaluated as an AI-assisted website planning/build option.
  • Canva — used in the broader ebook and visual-content workflow.

Other models or tools such as Claude.ai, Gemini, Grok, Perplexity, Midjourney, Ollama, and LM Studio are not included as confirmed experience because the available project records do not verify their use.

Confirmed environments

Supplied Nous Portal promotional screen showing the Hermes Agent name, available subscription tiers and Nous Research's MIT licence. It is supplied artwork, not a record of a purchased plan or of the installed agent version.
ComponentDetails
WorkstationWindows 11; host for Hermes, the Orchestrator foundation, Windows service, and named-pipe IPC
Cloud providerTencent Cloud VPS
Instance typeStarter Type
VPS operating systemUbuntu 24.04.4 LTS
CPU2 vCPUs
Memory2 GB RAM
Storage40 GB Cloud SSD
Network transfer512 GB/month
Network bandwidth20 Mbps
Hermes Agentv0.21.3 at the last recorded check
Installation methodGit
Python environmentPython 3.12.3
Additional toolsPlaywright Chromium, browser tools, Git, Python tooling
Service managementHermes Gateway system service

Languages and technical components

  • Python
  • SQLite
  • pywin32 for Windows services and named-pipe inter-process communication
  • Unit testing
  • Markdown and YAML knowledge records
  • Google Sheets as a registry component in the design
  • Windows security controls, including access-control lists (ACLs) and SeServiceLogonRight
  • Hermes custom skills, scheduled jobs, browser tools, and agent delegation

What This Work Demonstrates

My goal with AI is to create an extension of myself—a second brain that helps me think, organise ideas, solve problems, and execute my vision more efficiently. I do not put 100% of my trust in AI, nor do I let it think or make decisions on my behalf without scrutiny.

I remain the brain behind the work: I provide the ideas, direction, strategy, and judgment, while AI helps me turn my thoughts into action.

To me, AI is not a replacement for my intelligence or creativity; it is a tool that amplifies my capabilities, expands what I can achieve, and allows me to work more effectively while remaining in control.

The next conversation

Let’s connect the dots.

Project opportunities. Growth challenges. Career conversations.

Get in touch