Skip to main content
Emil Ingemar Karlsson — Agentic Engineer Stockholm

Emil Ingemar Karlsson

AI can build almost anything. What decides outcomes now is building the right thing for the right user, at the right price.

The principle is the same as before: fit matters more than volume. I connect data, models, tools, and workflows into automations that run in production with cost discipline and user value.

Working on right now

  1. Now

    Cost & guardrails

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

  2. Now

    Agent access & logging

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

  3. Now

    Memory & the right model

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

  1. 2026

    Agent pipeline after Paperclip

    Built a leaner pipeline from those learnings: GitHub as source of truth, a handshake gate between planning and execution, and OpenRouter model routing for fast-vs-hard tasks with cost control.

  2. 2026

    Autonomous organizations

    Ran automatic organizations in Paperclip — agents with roles, kanban, and approval gates — self-hosted on Coolify and Hetzner.

  3. Spring 2026

    Analytics platform

    Built a full analytics platform with Claude and Codex — Streamlit on the front, Databricks behind it, and the solutions running inside Databricks.

    Analytics platform case

  4. 2025–26

    Agent stacks

    Production agent stacks on Coolify and Hetzner: LangGraph, LiteLLM routing (Groq for cheap fast work, Claude for hard work), scoped tools, memory, and tracing.

    Production agent platform

  1. 2025–26

    Evals & approvals

    Quality gates and human approval before anything shipped — review, explicit stop-on-failure, and no publish or spend without a gate.

  2. 2025

    Data jobs into the platforms

    Started connecting data engineering work straight into Databricks, BigQuery, and Azure — making those jobs faster to run and cheaper to operate.

  3. 2025

    Building MCP

    Started building MCP servers, not only using them, so agents could work against scoped data and workflows.

  4. 2025

    MCP

    Connected the editor to live context — analytics, research, strategy, and tools — through Model Context Protocol.

    MCP development workflow

  1. 2025

    Cursor

    Started using Cursor. The model could read the repo, use local context, implement, and ship from the editor.

    See usage on my Cursor profile

  2. 2025

    Chat with data & retrieval

    Chat-with-data with retrieval: natural-language questions against live data, embeddings, and scoped context — plus n8n automations around collection, transformation, and delivery.

    Conversational analytics prototype

  3. Jul 2024

    First production AI product

    Built pipelines that translated to all global languages via the OpenAI API, using GPT-4o mini shortly after it launched on 18 July 2024. This was my first real production AI product.

  4. 2023–24

    Automation

    Wired models into n8n and Zapier so prompts became reusable flows: collect, transform, trigger, feed back.

  1. 2023–25

    NLP & ML on Databricks

    On the governed platform: forecasting, NLP across 1.46M+ support cases, and machine learning — applied AI on real enterprise data, not only chat.

    Enterprise data platform

  2. 2023–24

    Vibe coding

    The loop was ask, copy, paste, test, refine. It became building full sites and applications in conversation.

  3. 2023

    Claude for SQL

    Anthropic released Claude on 14 March 2023. I started using it for SQL — it became my default for data work.

  4. 30 Nov 2022

    ChatGPT

    OpenAI released ChatGPT. I started using it immediately for analysis, writing, and code.

  1. 2019–22

    Automation before GPT

    A few years of web scraping, Zapier, and RPA — automating collection and processes before language models entered the work.