fdw.expert
Floris de Wit · neverdecel

The last generationthat wrote codeby hand.

I type and read more than ever, just not in syntax. The interface between people and software has never been this fast. I'm trying to catch a glimpse of the future, but we're all going to live it.

Independent, based in the Netherlands. Working on agents and AI in production. Open to building with teams and researchers who are serious about it.

about

I was always the lazy engineer: anything I had to do twice, I never did by hand again. That kept running into limits: hands, time, logs nobody read. Those limits are falling away now, and not neatly. A lot breaks, and even more gets made. That is what I watch, build in and write about.

expertise, bottom to top

layer 1The foundation. Cloud platforms and everything as code, the base the rest runs on.For example: a full AI platform defined in Terraform and Bicep, rebuilt from a pipeline instead of by hand.

layer 2AI in production. From prototype to a system that runs and keeps running.For example: every model change goes through evaluation and an automated release, with observability on what it costs and returns.

layer 3AI that sticks. People, governance, security and cost, so it gets used and stays in use.For example: agents that act on behalf of the signed-in user with that user's own permissions, and every call in an audit trail.

layer 4Building with agents. Agents that do real work, with the tooling, evaluation and workspace around them.For example: coding agents with real repository and tool access over MCP, shared by a team instead of one person's laptop.

stack, per layer

layer 1Platform as codeAzure · Kubernetes (AKS) · Terraform · Bicep · GitHub Actions · Grafana · Datadog
layer 2Models in productionAzure AI Foundry · AI Search · evaluation · observability · automated releases
layer 3Access, traceability, costOAuth on-behalf-of · access control · audit trail · cost monitoring · enablement
layer 4Agents with real accessClaude Code · Codex · OpenCode · MCP · agentic workspaces · shared agents in teams

research

Behavioural research in sim racing: what a haptic signal in the brake pedal does to how people brake.

codeBrakeCuein developmentA haptic cue in the brake pedal that tells you how close you are to the grip limit. Tuned to how people perceive vibration, not to realism. Learns each car's braking limit per speed band from telemetry.

more on the research →want to test it? →

github.com/neverdecel

onbehalfOne shared AI harness setup for a team on one host. Each user keeps their own identity and credentials.Bashfdw.expertThis site. Astro, Markdown, Docker.Astro

publications01

21 Oct 2026Who is steering the conversation?No single party decides what language models tell us. There are five separate mechanisms, and any one of them is enough on its own.

Write me.

See it the same way, or not at all? Either makes for a good conversation.