AI journey

Three years of hands-on AI, from my first ChatGPT prompts in 2023 to leading my team's first agentic AI rollout. I learn by building things myself, then bring what works to the platform I run.

Jump to 2023 2024 2025 2026 Next

The short version

  • At work · Oct 2026

    Leading my team's first agentic AI rollout, governance first. Now in testing.

    Read the case study
  • Every day · since Oct 2026

    My own AI assistant over my notes, with guardrails so it can draft but never send.

    See how it's built
  • In public · Nov 2026

    retrieval-scoreboard: measuring which search method finds the right IT article.

    View on GitHub

2023 Learning to ask

  1. Early adopter

    • Used ChatGPT and Bard (now Gemini) every day, at work and on personal projects.
    • Learned to structure prompts: role, context, constraints and examples.
    • Generated images in Midjourney and DALL·E.
    • ChatGPT
    • Bard / Gemini
    • Midjourney
    • DALL·E

    What it showsHands-on from the first wave.

2024 Looking under the hood

  1. Running AI on my own machine

    • Ran local and small language models.
    • Built image workflows with Stable Diffusion in node-based ComfyUI.
    • Local LLMs
    • Small language models
    • Stable Diffusion
    • ComfyUI

    What it showsKnows the trade-offs between cost, privacy and control.

2025 Making AI do things

  1. Chatbots that know my documents

    • Built chat agents grounded in my own documents, using Botpress's knowledge base.
    • My first hands-on retrieval-augmented generation (RAG).
    • Botpress
    • RAG

    What it showsRetrieval fundamentals.

  2. Agents, tools and MCP

    • Built AI agents with tool calling and memory in n8n.
    • Built automations in Zapier.
    • Connected Claude to tools and data through MCP servers.
    • n8n
    • Zapier
    • MCP
    • Claude

    What it showsIntegration thinking, the same problem as enterprise workflow.

  3. Building with AI, not just prompting

    • Used Claude as a coding partner for scripts and small personal apps.
    • Claude

    What it showsShips working software.

  4. onward

    Building for my family

    • Small AI apps in Google AI Studio: a children's story creator and an anime image generator.
    • Illustrated storybooks starring my kids, made with Midjourney, Nano Banana and Canva. Kept private.
    • Short AI videos and ad films: stills in Nano Banana, motion in Kling, through Higgsfield.
    • Google AI Studio
    • Nano Banana
    • Kling
    • Higgsfield
    • Midjourney
    • Canva

    What it showsShips small, useful things fast.

2026 Doing it properly

  1. My own AI stack

    • Self-hosted n8n and local LLMs through Ollama.
    • PostgreSQL with pgvector for vector search.
    • n8n
    • Ollama
    • PostgreSQL
    • pgvector

    What it showsCan run AI on my own infrastructure.

  2. Independent RAG proof of concept

    • Built a retrieval-augmented generation proof of concept with Claude, Ollama and Pydantic AI.
    • Synthetic data only.
    • Claude
    • Ollama
    • Pydantic AI

    What it showsProves the idea before the build.

  3. At work

    Leading my team's first agentic AI rollout

    • Agentic workflows and a RAG knowledge base on ServiceNow.
    • Governance first: approved sources only, with hybrid search.
    • Now in testing.
    Read the case study

    What it showsDelivers AI safely at work.

  4. In daily use

    A second brain I actually use

    • My own AI assistant over my Obsidian notes, with hybrid search (meaning plus keyword).
    • A daily briefing, with read-only access to my email and calendar.
    • Guardrails: it can draft, but never send, post or delete. Built test-first.
    • Claude
    • Obsidian
    • Hybrid search

    What it showsGovernance by design, the same rules I apply to AI at work.

  5. In public

    retrieval-scoreboard

    • A synthetic 200-article IT knowledge base, a 30-question test set and a scoring harness.
    • Next: meaning vs keyword vs hybrid search, scored with real numbers.
    View on GitHub
    • Python
    • Search evaluation
    • Test-driven

    What it showsMeasures, doesn't eyeball.

Next Where I'm heading

  1. Exploring

    What I'm exploring

    • Multi-agent architecture.
    • More standalone personal apps.
    • retrieval-scoreboard Levels 2 to 7: hybrid search, reranking, cited answers and observability.

    What it showsKeeps learning in public.

Bringing AI to your ServiceNow platform?

I run ServiceNow platforms and bring AI into them carefully, with governance, tests and real measurements. If your team needs that, let's talk.