Selected work

What I have
built.

Different clients, one aim: take the repetitive, manual work off people so they get their time back. A production AI recruitment stack, a second brain that rewrites itself, and the ventures, builds and studies behind them.

01 · 2026Flagship

Neuron by Thalamus

I built one place that holds everything a company knows and does the busywork on top of it, so people stop feeding a dozen tools and get to be people again.

Founder and lead engineer

Next.js 15SupabasePostgreSQL + RLSTypeScriptVercelMCP integrations

Companies do not lose their knowledge because people are careless. They lose it because it lives in intuition nobody knows how to write down, scattered across a dozen tools nobody ever fully learns. Neuron fixes both. Each person builds their brain just by answering questions, so the things they only knew by feel finally get captured, in their own words. It plugs into the tools they already live in, Gmail, calendar, the CRM, Notion, and pulls the knowledge out of them into one simple screen anyone can ask, instead of ten logins and ten ways of working. And it does not just remember, it acts: drafting in your voice, sending the weekly update, handling the routine machine work. That is the whole point. Machines should do the machine work, so people get to do the human part. Thalamus is the brain underneath. Neuron is what puts it inside a company.

  • People build their brain just by answering questions, so the know-how they only had by instinct and never wrote down finally gets captured, in their own words
  • One simple screen holds what the whole company knows, so people stop hopping between a dozen logins and ten different ways of working just to find one answer
  • It plugs into the tools they already use, Gmail, calendar, the CRM, Notion, and pulls the knowledge out of them into the brain instead of leaving it stranded in each app
  • It does not just remember, it does the work: drafting in your voice, sending the weekly update, taking the routine tasks off people so they are free for the part only a person can do
  • Everything is sealed per company by security I broke into myself until nothing leaked, private notes stay private, and what people choose to share becomes one brain the whole team can ask
Radical Recruitment AI Stack screenshot
02 · 2026Flagship

Radical Recruitment AI Stack

I built an AI that scores people on who they are, not what their CV says.

Head of Technology & co-founder

Next.js 15 / 16Supabase + pgvectorOpenAI / Gemini embeddingsAstro 6Vitest / PlaywrightVercel
radicalrecruitment.ai

A solo-built, production recruitment platform where OpenAI embeddings and pgvector do the semantic matching, a proprietary seven-component engine scores candidates on four human dimensions, and a transparent gate decides who enters the pool, all behind a CI-enforced security suite.

  • Semantic matching: 768-dimension embeddings (OpenAI or Gemini, swappable behind one provider abstraction) plus pgvector cosine search in Postgres
  • The APAC engine: seven weighted components that sum to one, role-aware so the weights shift per detected role type, and every score explainable
  • De Poort: a learning-then-active gate that decides pool admission, with every rejection human-checked
  • Production security: row-level security on every table, a 57-assertion pentest regression set and an RLS verifier, all wired into a CI gate (lint, typecheck, security, build)
  • 165 automated tests in the CRM plus Playwright end-to-end, and one person owning the AI, the data model, two apps, the public website and the security
03 · 2026Flagship

Thalamus

I built an autonomous second brain that works like a real one, digitally, and lifts the manual work and mental load off me.

Sole architect

Claude CodeMCPObsidian vaultMarkdown knowledge baseSkills

Thalamus is my second brain. An autonomous system that works the way a real brain does, digitally: it holds my whole world in a structured knowledge base, reasons over it to draft and act in my voice, reads my live inbox and calendar, and rewrites its own instructions when it gets something wrong. The name is personal. The thalamus is the part of my own brain a haemorrhage struck, the part I fought my way back from. So I built a new one in code, to carry the work and the load I no longer want to carry by hand.

  • A second-brain architecture from scratch, the Four C's: Context (who I am and my work), Connections (live data), Capabilities (skills), Cadence (its rhythm), every decision reasoned through what it knows about me
  • Retrieval against a hand-curated markdown knowledge base kept as an Obsidian vault, version-controllable and human-readable, instead of a black-box vector database
  • Three live MCP integrations: Microsoft Graph (Outlook), Google Workspace (Gmail, Calendar, Drive) and Notion, each under strict scoped permissions so it can read and draft but never sends or posts on its own
  • Custom model-invokable skills, including a tone engine that captures two distinct writing registers from my real sent email, a job-application pipeline, and a task board
  • A genuine self-improvement loop: every miss fixes the source file or skill and is logged so it cannot recur, with scoped local tokens and explicit allow and deny lists
04 · 2026Flagship

Market Signal AI

The model that knows when to shut up.

Founder & sole builder

Pythonscikit-learn / LightGBM / CatBoostSHAPDutch open data (CBS / KVK)FastAPINext.js 16

A machine-learning platform that makes data science accessible to any company, not only enterprises with their own data teams. It learns from a company's CRM to find lookalike customers and scans Dutch open market data to surface untapped sectors and regions, runs its own model contest, calibrates the winner so the probabilities are real, explains every score, and refuses to predict when a held-out AUC gate proves the data has no signal.

  • Two-way intelligence: an inside-out engine that learns from your CRM to find lookalike customers, and an outside-in engine that ranks Dutch sectors and regions from free public data (CBS, KVK, OpenStreetMap, TenderNed)
  • An auto model bake-off across LightGBM, RandomForest, CatBoost and a stacking ensemble, chosen on cross-validated top-decile precision with no human in the loop
  • A 0.78 to 0.97 AUC honesty gate plus a supervised to lookalike to market-signal fallback ladder, so weak data is never dressed up as a prediction
  • Probability calibration (auto isotonic versus sigmoid by Brier score) and per-lead SHAP explanations in plain Dutch and English
  • Validated on real third-party data: AUC 0.93 and 0.80 where signal exists, and a correct refusal at 0.51 where it does not, the product's whole thesis made true
  • Scores companies, sectors and regions, never individual people, a deliberate EU AI Act positioning that keeps it broad and safe to sell across industries

The rest of the constellation

Every star here was earned.

Servicepaspoort Website & Blog Engine screenshot
2026

Servicepaspoort Website & Blog Engine

Digital Transformation Specialist

Making an LLM physically unable to lie: every sentence traces to a real source, and it cannot publish an uncited fact or a medical dose.

I led the digital transformation at Servicepaspoort, a Dutch senior-care brand. I built their new website, the live servicepaspoort.nl, with its content architecture and a writing system tuned for an older audience, and the guardrailed, RAG-based AI engine that fills it, built around one hard client rule: the AI may invent nothing. Built end to end, solo.

  • Designed and built the new servicepaspoort.nl end to end: the page architecture (a news bento grid plus a five-theme content carousel), the sections (Ons aanbod, Club Vitaal, Nobilis, member stories), the contact flow and the webshop
  • Authored its brand voice and content strategy: a babyboomer audience model and a copywriting style guide (aspirational, quality-led, regie and empowerment, never patronising), applying Deloitte 'show, do not tell' UX so emotional connection turns into membership
The full story
  • Mandatory retrieval, not open generation: a curated Postgres library of around 450 source URLs in roughly 1,400 paragraph chunks across 31 Dutch care topics, vetted against around 28 official Dutch health authorities
  • A six-layer safety architecture, most of it deterministic code: input guard, regex citation validator, AI fact-checker, live source-verifier, output guard, and strict retrieval that refuses if fewer than three sources match
  • Every factual sentence is auto-checked for a citation to an exact source page; the output guard masks dosing, hedges reader-directed diagnoses, and redirects treatment advice before export
  • A multi-provider LLM layer with streaming fallback across Gemini 2.5 Flash, Claude Sonnet 4, Groq (Llama 3.3 70B) and DeepSeek, so a provider failure switches instead of dead-streaming
  • A 100-point GEO/SEO/AEO scoring engine and a human-on-the-loop reviewer panel, on Next.js 16 and React 19
Website + CMSNext.js 16React 19Supabase (FTS + pgvector)RAG
0-layer
safety pipeline
0
vetted source chunks
0
LLM providers, fallback
We Know People screenshot
2025

We Know People

MarTech / AI Specialist

Sixteen months as the in-house AI specialist, turning LLMs into teammates and automating the operation around them.

As the in-house MarTech and AI specialist at a Dutch recruitment agency, I built and deployed the AI agents, automations and LLM integrations that ran the marketing and recruitment operation: an autonomous vacancy radar, GDPR-proof contact automation, an AI blog generator, and the development playbook the company builds its tools by.

  • Sixteen months owning AI, automation, HubSpot and BI for a firm with a 30,000+ professional network
  • Built an autonomous vacancy radar in n8n: it watches former clients' career pages, strips and cleans the raw HTML with my own crawler, then orchestrates three LLMs (OpenAI, OpenRouter and Google Gemini via LangChain) to extract structured vacancy data and draft a personalised recruiter email in Gmail, posting to Microsoft Teams and logging to Google Sheets
The full story
  • Built GDPR-proof marketing automation: a HubSpot and Otys API sync with automatic opt-out handling and dynamic cohort segmentation, so campaigns stayed targeted and compliant with no manual list management
  • Shipped an AI blog generator: forced Gemini 2.5 Flash into typed JSON output (SEO analysis, full article and FAQ), then chained Imagen 3 for a text-free illustration and alt-text per section, with one-click export to Markdown or a Docs-ready HTML file
  • Authored the company's digital-tools development playbook, a Define, Design, Execute, Learn process built on the Double Diamond and Agile sprints, so every new internal tool shipped to a repeatable standard
n8nLangChainOpenAI / Gemini / OpenRouterHubSpot + Otys APIImagen 3
0
months owning AI & automation
0
LLMs orchestrated in one pipeline
0+
professional network
EIPSI / Evidence for Teaching screenshot
2021

EIPSI / Evidence for Teaching

EdTech & Digital Transformation Innovator

Designing evidence-informed teaching for Europe's hardest schools, online and off.

On a 371,181 euro EU Erasmus+ project across five countries, I ran the full design-thinking process to build the EIPSI Hub and an offline-first teacher toolkit for low-connectivity schools.

  • Part of a publicly credited EU Erasmus+ consortium (371,181 euro, six universities across five countries)
  • On a six-person team, reframed a vague social brief into a sharp design question, then ran research, ideation, prototyping and user testing
The full story
  • Co-built the EIPSI Hub (evidenceforteaching.org), the project's open digital platform for teachers, counsellors, inspectors and administrators
  • Designed an offline-first printable toolkit (a card game, infographic and school materials) for schools with poor connectivity, the key insight of the project
  • Tied to the UN goals of Quality Education and Reduced Inequalities, with a clear equity mission
Design thinkingUX / UIPrototypingFigma
0
EU Erasmus+ project
0
countries
0
universities
Koninklijke Marinedagen App screenshot
2022

Koninklijke Marinedagen App

Project manager, 15-person team

I shipped the Royal Navy's official event app for 100,000 visitors, on a deadline that could not move.

When the Royal Netherlands Navy needed an app for its first public Navy Days in five years, I ran the 15-person team at DTT that delivered it on time, with augmented reality of the fleet, QR ship info, and a live event map.

  • Led a 15-person full-stack team, translating client goals into shipped features against a fixed, unmovable event date
  • Around 100,000 visitors used it across two days at a national naval event
The full story
  • Augmented reality of the fleet plus QR scanning for ship-level information: real interactive tech, not a brochure app
  • Five-star rating and a written testimonial from the Royal Netherlands Navy, calling DTT a reliable partner
Project managementNative iOS + AndroidARQR
0
visitors in two days
0
person team I led
0-star
Navy rating
Inholland Student Initiative Fund screenshot
2020

Inholland Student Initiative Fund

Founder & project leader

I founded a 10,000 euro per idea student fund. 200+ initiatives later, it still runs without me.

I turned an institutional budget students could not reach into a student-run fund with a two-week path from idea to funding, then built the team and the operating model to keep it alive.

  • Founded and launched it in December 2020, designing the operating model, the application-to-decision process, and the first student-worker team
  • Up to 10,000 euro per idea, with a supervisor meeting within two weeks of applying
The full story
  • 200+ initiatives funded and counting, from short films and study trips to robotics builds and a student radio station
  • Still running in 2026 with around seven student employees and two staff coordinators, long after my own tenure
  • Done while chairing the Programme Committee and as the sole student member of the Advisory Group for roughly 26,500 students
0 to 1Operating modelGovernanceBudget stewardship
0+
initiatives funded
0
per idea
0
students represented
2026
2026

Haarlem Home Sale Campaign

Go-to-market strategy & copy

I ran an apartment sale like a go-to-market campaign, not just a property listing.

To reach beyond the standard Dutch buyer pool, I vetted the highest-reach expat and local housing communities by size and rules, then wrote owner-voice outreach and a full listing in three languages, with a Schiphol-commute angle and an anti-spam posting plan, to put a Haarlem new-build in front of the right international buyers.

  • Vetted each channel live and ranked by reach and relevance, from a 56,900-member expat group to active Haarlem-local and Schiphol-area communities, flagging public versus join-first groups
  • Wrote a short owner-voice note plus a full listing (price, 91 m2, A+++ energy, EV-charging parking, transport links) so it read like owners, not an ad
The full story
  • Built audience-specific angles: a commute opening line for Schiphol-area groups and a Spanish-language variant for Spanish-speaking buyers
  • Planned an anti-spam rollout, a few groups per day split between the two owners, to stay credible and avoid platform flags
Go-to-marketChannel researchAudience targetingCopywritingDutch / English / Spanish
0
Haarlem new-build listed
0+
targeted channels vetted
0
languages of the listing

Want to see who built all this? Read the story.