I’m Floris Vossebeld. I work on applied AI in regulated enterprises (mostly financial services) at Microsoft. The job sits in an awkward, interesting place: between the demo that impressed everyone in the room and the system that has to survive real data, permissions, latency, cost, and someone owning it on Monday.
That gap is what this site is about.
A lot of AI looks convincing right up until it touches any of that. I spend my time close to where it does, with banks and insurers and the constraints they can’t wish away. So I’m less interested in whether a model is impressive than in what it takes to make one a governable, useful part of a system people actually depend on.
Where I come from
My background is in computer science and data science, and the thread running through it is tool use, feedback loops, and systems that can correct themselves. That’s the lens I brought to my master’s thesis: using reinforcement learning to teach an agent to walk a knowledge graph, use tools, and recover from its own wrong turns instead of answering in one shot. It’s still how I look at most things. The route here:
- 2019–2022 BSc Business Information Technology · University of Twente Computer science with a business-systems bent.
- 2023–2025 MSc Computer Science, Data Science · University of Twente Magna cum laude. Thesis graded 9.5: reinforcement learning to make an LLM build and refine SPARQL queries over a knowledge graph.
- 2024–2025 Co-founder · ChatIT Trained teams to use tools like ChatGPT and advised leaders on where generative AI actually cuts cost.
- 2025 Microsoft internship → ISWC 2025, Nara The thesis grew out of a Microsoft Netherlands internship. I presented Learning to Refine at the RAGE-KG workshop: 49.7% on LC-QuAD 2.0, up 17.5 points on the prior iterative baseline.
- Aug 2025 – now Solution Engineer, Cloud & AI Apps (Financial Services) · Microsoft Hands-on with the largest banks and insurers, moving GenAI from proof-of-concept to production.
What I write here
This is a notebook, not a brand. Three kinds of thing end up here:
- Essays
- The longer arguments, like the case that AI is shifting from interface to infrastructure.
- Technical sketches
- One diagram, one idea: retrieval beyond vector search, federated memory, approval flows (the architectures I keep redrawing).
- Field notes
- Occasional and less technical: what demos teach about clarity, and the courage it takes to ask the obvious question.
The current spine is From chatbots to system operators. Mostly the systems; sometimes the human parts of doing the work. If there’s one bet under all of it, that’s it. The hard problems are the boring-sounding ones: trust, ownership, evaluation, scope.
Find me elsewhere
- GitHub: @FVossebeld
- LinkedIn: floris-vossebeld
How this site works
This is a public, version-controlled wiki. I curate the sources and stay the editor-in-chief; an AI agent does the cross-referencing and maintenance. The source is fully open: view or fork it on GitHub. See how-this-works for the full picture.