Skip to main content
Emil Ingemar Karlsson

Emil Ingemar Karlsson

Engineer and founder building across the stack: data platforms, AI automation and agents — and the business conversations that decide what's worth building.

AI changed how fast things get built. It didn't change what's worth building.

I've worked through every layer of that shift — data platforms and ML in production, AI automation and agents that run real workflows — always alongside the business side. The steps below are that path. The longer write-up is From BI reports to agents.

The AI journey

Working on right now

  1. Now

    Business

    Cost as a design constraint

    Failure-priced ROI and spend policies per agent role. The business case stays part of the build.

    Failure-priced ROI

  2. Now

    Agents

    Cost & guardrails

    Cost control and guardrails so autonomy stays cheap, scoped, and accountable.

    Guardrails in practice

  3. Now

    Agents

    Agent access & logging

    Who agents can reach, and a log of what they did — access control and tracing as a first-class area.

    Company OS

  4. Now

    Agents

    Memory & the right model

    Building learning and memory so context accumulates, and routing so the right model is used for each job.

    Model routing notes

  1. 2024–now

    Data

    Hockey analytics platform

    SHL, NHL and more: scrapers, DuckDB/MotherDuck, and an automated publishing loop — same data-craft discipline applied to the game.

    Hockey Analytics case

  2. Spring 2026

    ML & GenAI

    AI analytics platform

    Streamlit on Databricks, with GPT-4o turning questions into validated SQL and narrative answers, built with Claude and Codex.

    AI analytics platform

  3. 2026

    Agents

    Agent organisations, then a leaner pipeline

    Ran agent organisations with roles, kanban and approval gates. The lessons became a leaner pipeline: GitHub as source of truth, a planning → execution handshake, and cost-aware routing.

    Company OS learnings

  4. 2026

    Data

    LINHAC 2026

    Conference notes on analytics as club infrastructure — pipelines and process over one-off dashboards.

    LINHAC report

  1. 2026

    Data

    SHL cores beat coaches

    52 team-seasons: returning-core share correlates with points%; coaching changes barely move the club level.

    Method note

  2. 2025–26

    Agents

    Agent stacks

    LangGraph, LiteLLM routing (fast models for cheap work, stronger models for hard work), scoped tools, memory and tracing, self-hosted.

    Production agent platform

  3. 2025

    ML & GenAI

    Chat with data & retrieval

    Natural-language questions against live data with embeddings and scoped context.

    AI analytics platform

  4. 2025

    Agents

    Cursor

    The model reads the repo, uses local context, implements and ships from the editor.

    Cursor profile

  1. 2025

    Data

    Data jobs into the platforms

    Data engineering work connected straight into Databricks, BigQuery and Azure, so it runs faster and costs less to operate.

    Analytics platform case

  2. 2025–26

    Agents

    Evals & approvals

    Quality gates and human approval before anything ships or spends.

    Eval fixture notes

  3. 2025

    Agents

    MCP, then building MCP

    Connected the editor to analytics, research and tools through MCP, then built MCP servers so agents could work against scoped data.

    MCP development workflow

  4. 2025

    Automation

    Slack to live site in 30 seconds

    A structured message becomes a deployed article through n8n, GitHub and Netlify.

    AI automation in production

  1. Jul 2024

    Automation

    First production AI product

    Translation pipelines to all global languages via the OpenAI API, using GPT-4o mini shortly after its launch.

    AI automation in production

  2. 2023

    ML & GenAI

    Claude for SQL

    Became my default for data work soon after its release in March 2023.

  3. 2023–26

    Data

    Governed lakehouse

    100 production jobs and 600+ governed Unity Catalog tables on Azure Databricks, with a Bronze → Silver → Gold model across 20+ source systems.

    Enterprise data platform

  4. 2023–25

    ML & GenAI

    NLP & ML on the platform

    Topic modelling across 1.46M+ support cases, churn and demand-forecasting models, all tracked in MLflow.

    Enterprise data platform

  1. 2023–24

    Automation

    Prompts become flows

    Models wired into n8n and Zapier, so prompts became reusable flows.

    Feedback agent case

  2. 2023–24

    Automation

    Vibe coding

    Ask, paste, test, refine: building full sites and applications in conversation.

  3. 30 Nov 2022

    ML & GenAI

    ChatGPT

    Used from day one for analysis, writing and code.

  4. 2021–23

    Business

    One evidence base, 35 markets

    Product analytics, 1,470 survey responses and 30 customer visits guided platform investment across 35 markets.

    B2B decision case

  1. 2019–22

    Data

    Automation before GPT

    Web scraping, Zapier and RPA: automating collection and processes before language models entered the work.

  2. 2019–21

    Data

    Client BI dashboards

    As a BI consultant at XLENT: scoped and delivered Power BI across client assignments. GDPR and security work made trust part of every report.

  3. 2019

    Data

    Power BI Planner

    Delivery dashboard over Microsoft Planner exports for standups and prioritisation across many boards — agile follow-up without opening every plan.

  4. 2019

    Data

    Support analytics dashboard

    Power BI + SQL on ServiceNow: from reactive ticket closing to pattern detection, early intervention and a clearer view of support performance.

    Related ops case

  1. 2018

    Data

    SLA reporting → ops dashboard

    At TietoEVRY: ServiceNow history into SQL datasets and a Power BI dashboard for recurring incidents, resolution trends and earlier intervention — not just retrospective SLAs.

    Operations analytics case

  2. 2017

    Data

    Food-waste marketplace MVP

    Lean Django marketplace in Jönköping: restaurants published surplus food in a short window; customers browsed and collected take-away.

    Early products

  3. 2016

    Data

    Digital ticketing admin

    Django event and ticket admin with sales tracking for small organisers. Early full-stack product work: Sketch → Bootstrap → Django.

    Early products

  4. 2016–17

    Business

    Early product builds

    Digital ticketing and a food-waste marketplace MVP: cheap to run, hard to change behaviour. Product lessons that still shape what I treat as worth building.

    Early products

  1. 2016

    Business

    QuizFlow — killed in prototype

    Sketch and InVision quizzes for engagement and lead-gen. User tests showed no real demand — stopped before build. Knowing what not to ship.

  2. 2015–16

    Data

    Product analytics from day one

    TUVA shipped with Mixpanel and Google Analytics. Behaviour data shaped product decisions from launch — analytics was never a later add-on.

    TUVA case study

  3. 2014–16

    Business

    Founding TUVA

    From Swift prototype to 10,000 users in three weeks, with behavioural data from launch.

    TUVA case study

About

I build across the stack, and the breadth is the point: data platforms, AI automation and agents, and the product decisions that decide what gets built. Founded TUVA; ran governed lakehouse and ML in production; operate agent systems through The Unnamed Roads. Based in Stockholm.