AI Deployment Strategist — responsible real-world AI implementation.
I translate organisational problems into AI workflows that can actually be adopted, governed and used. My work sits where AI becomes operational: workflow design, deployment decisions, and the privacy and review boundaries that make implementation credible in public-sector and organisational settings.
- Turn operational needs into deployable workflows — not demos. Bounded scope, real users, documented trade-offs.
- Build and operate the platform I run them on — routing, delegation, memory and cost accounting, extended as needs appear and retired when they do not earn their keep.
- Keep the boundaries explicit — what a system must never receive, who decides what gets published, and what is deliberately left out of a public writeup.
| Repository | What it shows |
|---|---|
| ndt-ai-workflows | Privacy-aware AI assistant architecture inside a Microsoft 365 environment — adoption, support workflow and governance in a real organisation |
| xr-fundraising-reporting | Multi-source reporting pipeline and executive decision support, with a privacy-preserving donor model |
| bvnw-news-assistant | A bounded editorial agent with human approval as the publishing gate |
| hermes-tooling | Extending my own agent platform — adopted tools measured rather than assumed, vendor claims separated from results, and negative findings published |
| wikipedia-editorial-agent | Goals and governance for AI-assisted Dutch Wikipedia work — corrective search, neutral prose, human authority at the publication gate |
- Adopt and measure before building. The measurement is usually the more useful finding, and it is often less flattering than the marketing.
- Separate claims from evidence. A vendor benchmark is a hypothesis until it reproduces on your own workload.
- Publish the negatives. A hypothesis that did not hold is worth recording — and a component left running without a verdict is a liability.
- A person decides anything consequential. Agents propose; humans publish.
- State the boundary up front — what the system must never be handed, written before the integration rather than after an incident.
Not pinned, but part of the record:
- happy-bee — Raspberry Pi-assisted micro-drone experiment: FPV video ingest, face recognition and PPM control. An older hands-on hardware/computer-vision project, kept readable.
- citizens-assembly-knowledge-base — a modular site with a chatbot answering questions about citizens' assemblies; early civic-tech exploration.
- italiaans-oefenapp — small tool for practising Italian verb conjugations.
Earlier work on this account also includes an IP-camera facial-recognition script (2018, a fork) and some private repositories used for client and pilot work.
Building tooling around my own agent platform and documenting the parts that were worth keeping. Wageningen, Netherlands.