I build the tools that make IT teams faster.

Why

IT operations work spans a lot of systems. Looking up a user's group memberships, checking a device's compliance state, pulling a figure from a report and reviewing a quarantined email each meant a different portal. The integrations remove that context switching.

How it works

Each integration is a separate MCP server exposing a focused set of tools — read operations by default, with write operations gated behind explicit confirmation. Authentication runs through Entra ID app registrations with least-privilege scopes, and privileged actions require an active PIM role.

Problems worth mentioning

• A third-party package wrote diagnostic output to stdout, which corrupted the protocol handshake. I patched it with a PowerShell script that runs as part of setup.
• Identity mismatches where a connector authenticated as the standard user rather than the elevated admin account, causing intermittent permission errors on privileged calls.
• Windows-specific path and runtime issues that don't appear when developing on macOS.

Rollout

Some servers run locally, others are hosted in Azure App Service. Each has a documented query playbook on Confluence so colleagues can use them without needing to understand the implementation.

Architecture

A Flask backend split into service modules — one per integration: ticketing, privileged access management, and documentation. The frontend is a single Jinja2 template with vanilla JavaScript and a dark theme, kept deliberately simple so there's no build step to maintain.

What the AI layer does

• Drafts replies from a short description, in the right tone and language
• Suggests a category and priority for incoming cases
• Summarizes long communication threads so you can pick up a case someone else started

Integration details

The ticketing API had some non-obvious semantics — message type rather than direction determines whether a note is an agent reply or a user message, and the sender ID is not the same as the requester ID. Getting those wrong produced threads that looked correct but attributed messages to the wrong person. Privileged role activation runs through Microsoft Graph directly from the tool.

My role

This sat outside my normal scope — a technical project alongside the internal development team, which meant learning their toolchain rather than bringing my own.

What I built

• Python scripts to locate and process the relevant files and reports feeding into AI analysis
• Automation of the data flow between systems, replacing manual steps
• Debugging of integrations where data arrived in the wrong shape or not at all

Environment

Azure DevOps for source control and pipelines, Data Factory for orchestration, Databricks for processing, and Power Apps on the front end. Documented the technical solution so it didn't depend on me being available.

Backend

Express serves the static frontend plus two endpoints: /api/status, which drives the live status bar at the top of the page, and /api/contact, which validates and stores form submissions.

Deployment

• Containerized with a small Alpine-based Docker image
• Hosted on Render, with automatic deploys on every push to main
• Custom domain configured through A and CNAME records, with HTTPS certificates issued automatically

Bilingual setup

English and Norwegian are separate pages sharing one stylesheet and one script. The script reads the lang attribute on the page and picks its strings accordingly, so there's no duplicated JavaScript and no client-side translation layer.