Brian Wagner builds the AI layer on top of the infrastructure he already runs

Now

Building agentic AI workflows with Claude Code and MCP
Internal platforms pulling data from 20+ systems into one store
Keeping LLM cost down with a human approving what matters

Learning

TerraformGitHub Actions CI/CDKubernetes (k3s)AzureOpenTelemetryEvalsRAGAgent patterns

Background

25 years of enterprise infrastructure operations
VMware, Citrix, Active Directory, Veeam, firewalls and M365
Security work: SIEM, endpoint protection, vulnerability management

About me

I connect AI to the systems that actually run the business

My leverage is everything around the model: deploying it, securing it, monitoring it, controlling its cost and wiring it into real systems. That is infrastructure work pointed at a new target. After 25 years I know what is worth asking across backup, virtualization, identity and security tools, because I know how they break.

I build internal platforms that pull data from many vendor systems into one operational store, then put an agentic layer on top: tools the model can call, routing to the right source, and a human approving anything that matters. I also run a full homelab, where I try things before they go anywhere important.

Selected results

Rounded figures from platforms I have built and run.

~40%

lower LLM cost per interaction, with throughput up at the same time

20+

vendor systems brought together into one operational data store

~20

tools defined for agentic routing across that stack

100x+

faster reporting queries after restructuring the data layer

300+

helpdesk tickets a year removed by automating one recurring notice

~80%

of routine helpdesk work traced to six repeatable streams, each a candidate for automation

Experience and focus areas

AI platform and agentic engineeringClaude Code, MCP, tool calling and routingNow

LLM cost and governanceContext management, model selection, human in the loopNow

Data and integrationETL pipelines, PostgreSQL, REST APIs across 20+ systemsDaily

Infrastructure operationsVMware, Citrix, Active Directory, backup, firewalls, M36525 years

Security and compliancePermission auditing, SIEM, endpoint security, vulnerability managementOngoing

Homelab and self-hostingDocker, Caddy, single sign-on, monitoring, pull-based deploysPersonal

My stack

Claude CodeMy main tool for building: agentic coding, skills and automation

MCP and tool callingWiring models to real systems with well-defined tools

Anthropic API and OllamaPrompt design, cost control and local models for bulk work

Python, FastAPI and HTMXFast, simple internal platforms and dashboards

ASP.NET Core and PostgreSQLOperational data stores and self-hosted apps

Docker, Caddy and LinuxContainers, reverse proxy and server administration

VMware, Citrix and VeeamVirtualization, virtual desktops and backup and recovery

Active Directory and Entra IDIdentity, Microsoft 365, Graph API and Exchange

Security toolingElastic SIEM, SentinelOne, Nessus, Netskope and FortiGate

Git and GitHubVersion control, releases and open-source projects

What I am learning next

A one-year plan, learned by running real AI workloads at home.

Oct to Dec 2026Terraform and CI/CD with GitHub ActionsNow

Jan to Mar 2027Kubernetes (k3s at home) and Azure basicsNext

Apr to Jun 2027Observability for AI: OpenTelemetry traces of agent runsPlanned

Jul to Sep 2027Capstone: an AI platform deployed with Terraform and traced end to endPlanned

Lone Herd

One experienced hand, a whole herd to move

This is the name I work under. It comes from the roundup: gathering scattered cattle into one herd before anyone can move them. I do the same with scattered IT data, with AI doing the riding.

Read the story

FAQ

Answers to common questions about my work

What are you working toward?

AI integration and internal AI platform engineering: building the plumbing that lets AI reach the systems that matter, and bringing people and systems together.

Are you a machine-learning engineer?

No. I do not train models. My leverage is everything around the model: deploying it, securing it, monitoring it, controlling its cost and wiring it into real systems.

How do you keep AI costs under control?

Context management, picking the right model for each job, and measuring cost against the work delivered. That approach cut LLM cost per interaction by roughly 40% while throughput went up.

How do you keep AI safe around production systems?

A human approves anything destructive, tools are read-only by default, and every tool definition documents its API contract and known failure modes. I snapshot before and after changes.

What do you use day to day?

Claude Code for building and automation, MCP for tool access, local Ollama models for bulk, low-stakes work, and Gemini for image and video generation.

Where can I see your work?

On GitHub, and in the Featured Works on the home page.