Magnus Lixenstrand

AI Engineer

LLMs in real business workflows

I automate away unnecessary manual work, these days with LLMs. Twelve years in sales taught me where work gets stuck, so I planned and built Nordic Metal Trade’s ERP system from scratch. Today I’m its product owner and sole developer.

About ten colleagues use the system every day, and it saves the team more than 20 hours a week.

See the ERP system Open CV (PDF) CV på svenska
Magnus Lixenstrand

An in-house ERP system for the whole workflow

Nordic Metal Trade sells steel and metals to industrial customers in Europe. The ERP system is tailored to how the company works.

  7. Accounting
A material enquiry: search hits among suppliers and past orders on the left, three material lines with dimensions, steel grade and weight in the middle, and the chosen supplier on the right
Material enquiry. One search goes through suppliers, past orders, quotes and products at the same time. The enquiry to the chosen suppliers is created in the same view.

The screenshots come from the test environment, and all data in them is made up. The system can only be reached through the company VPN, so you see screenshots here instead of a demo.

Starting point
Quotes were made in Excel, and the same details were typed in several times. There was no single place to see where a case stood.
The start
As a salesperson, I first automated quoting, moving it from Excel and VBA to a web flow connected to Fortnox. Today’s system grew out of it.
My role
Product owner, sole developer and project lead for the rollout: requirements, data model, development, testing, training and the switch from the old system.
Integrations
Fortnox, Shiplink, Unifaun/nShift and Microsoft 365. Orders, stock and purchasing flow into Fortnox without double entry, and freight is booked straight from the order.
AI assistant
Answers colleagues’ questions about orders, stock, materials and routines. Once the user confirms, it can also create customers and turn enquiries into quotes.
Technology
SvelteKit, TypeScript, FastAPI, Python, MySQL and Docker.

The agent workflow

I develop the ERP system with AI coding agents that take an issue all the way to a reviewed and merged PR, with me as the final check. More than 3,200 of the system’s 10,000-plus commits were made with agents.

If the tests fail, the agent fixes the problem itself, and I only take over if the same failure remains after three attempts. Another agent reviews the plan and the code, and the agents are spread across three computers based on how much quota and capacity is left.

Personal projects

Smart home

Home Assistant with 232 devices, 161 automations and 16 rooms, over Zigbee, Matter/Thread and MQTT. The automations connect lighting, heat pumps and Nord Pool electricity prices.

Hemma

A custom dashboard in SvelteKit and FastAPI that the family uses to run the home from phone and tablet, with push notifications and roles for family and guests. Its AI assistant proposes actions as buttons, and the app only carries out the one someone taps. Every proposal is checked against the house’s own lights, switches, scenes and scripts.

Household finances

A web app in SvelteKit, FastAPI and SQLite that pulls transactions from Toshl and tracks budget, savings goals and loans. It runs on a home server with a nightly reconciliation. Its own MCP server with 24 tools lets an AI assistant read the finances, check whether a purchase fits and recategorize purchases.

Majo

I’m continuing to develop Majo, an Android TV app that plays movies and shows from the user’s own Plex, Jellyfin and Emby servers. An LLM translates the subtitles while the movie plays. I’ve made a new TV interface with tactile focus. Agents build and publish new versions when I ask, and the app then offers them on the TV.
Hemma on a tablet. It runs locally and is connected to Home Assistant.

Get in touch

I’m looking for AI Engineer roles, in Jönköping or remote.

[email protected]