Work

Systems that
do the work.

Selected projects and solution blueprints: reference architectures for the problems I solve most often, from AI document intake to offline-first apps.

Selected projects

Recent work

Realesate agent
E-Commerce2025Featured

Madamerealty

Madamerealty

  • Nodejs
  • Nextjs
  • AI
View case study
Direct 247 ecommerece website

E-Commerce · 2026

Direct 247

An Ecommerce website for Business to Promote their Business Online

  • wordpress
  • WooCommerce
  • Rabo smart pay

Solution blueprints

Reference architectures I design and build.

Blueprints show how I approach common problems. They're patterns rather than client case studies; client names stay confidential unless a client has agreed to be named.

AI Automation

Document intake agent

Problem
Invoices, orders and forms arrive by email as PDFs and scans. Someone retypes them into the accounting or ERP system: slow, error-prone and boring.
Approach
An intake service watches the mailbox, an LLM extracts the fields into a strict schema, rules validate totals and suppliers, and a review screen handles uncertain cases before posting.

Flow

  1. Mailbox
  2. OCR + LLM extraction
  3. Validation rules
  4. Human review
  5. Accounting / ERP

OutcomeRoutine documents flow through automatically; people only look at exceptions.

  • Python
  • FastAPI
  • Claude / GPT
  • PostgreSQL
  • Exact Online / Moneybird API

AI Engineering

Support copilot with RAG

Problem
Support agents search through wikis, manuals and old tickets to answer the same questions, and new staff take months to get up to speed.
Approach
Company knowledge is indexed into a vector database. For each ticket the copilot retrieves relevant sources and drafts a cited answer that the agent approves or edits.

Flow

  1. Knowledge sources
  2. Chunking + embeddings
  3. pgvector search
  4. LLM draft with citations
  5. Agent approves

OutcomeConsistent, sourced answers and faster onboarding, with a human always in control.

  • TypeScript
  • Next.js
  • pgvector
  • Anthropic / OpenAI
  • Zendesk / Freshdesk API

Process Automation

Lead-to-CRM pipeline

Problem
Leads come in from the website, email and LinkedIn, and end up in inboxes and spreadsheets instead of the CRM. Follow-up depends on who noticed.
Approach
Every channel feeds one pipeline. Leads are de-duplicated, enriched with KvK and company data, scored by an LLM against your ideal customer profile, and created in the CRM with an owner and a task.

Flow

  1. Web forms & email
  2. De-duplication
  3. Enrichment (KvK)
  4. AI scoring
  5. HubSpot / Pipedrive + Slack

OutcomeEvery lead lands in the CRM with context and an owner, within minutes.

  • n8n
  • Node.js
  • HubSpot API
  • KvK API
  • Slack

APIs & Data

Operations dashboard

Problem
Management numbers live in five tools. Every Monday someone exports CSVs to build the same report by hand.
Approach
Scheduled pipelines pull data from the source APIs into a PostgreSQL warehouse, models define each KPI once, and a dashboard with alerts replaces the manual report.

Flow

  1. Source APIs
  2. Scheduled ETL
  3. PostgreSQL warehouse
  4. KPI models
  5. Dashboard + alerts

OutcomeOne trusted set of numbers, always up to date, with no Monday spreadsheet.

  • Python
  • dbt
  • PostgreSQL
  • Metabase / Power BI
  • Docker

Mobile & Desktop

Offline-first field app

Problem
Technicians fill in paper forms on site with no reliable signal, and the office retypes them days later.
Approach
A React Native app stores inspections locally, captures photos and signatures, and syncs to the backend when a connection returns. The office sees results in the web portal.

Flow

  1. Mobile app
  2. Local database
  3. Background sync
  4. API
  5. Web portal

OutcomeReports are complete and in the system the same day, with no double entry.

  • React Native
  • Expo
  • SQLite
  • Node.js
  • PostgreSQL

Web Platform · Live

This website

Problem
A bilingual business site that editors can manage without a developer, fast enough to rank, and memorable enough to stand out.
Approach
Next.js with a headless Payload CMS, English and Dutch on separate domains, structured data for search, a WebGPU fluid simulation behind the home page and live shader graphics on every page header.

Flow

  1. Payload CMS
  2. Next.js App Router
  3. next-intl (EN / NL)
  4. WebGPU (vgpu)
  5. Self-hosted with Nginx

OutcomeReal, and live: you're using it.

  • Next.js 16
  • Payload CMS 3
  • PostgreSQL
  • vgpu / WebGPU
  • Tailwind CSS

Toolbox

The stack behind the work

Frontend

  • Reactjs
  • Nextjs

Strategy

  • AI

Backend

  • Hono js
  • Nestjs
  • Nodejs

15+ projects

Delivered across AI, automation, web, mobile and data.

95% satisfaction

Client satisfaction across delivered projects.

End to end

From scoping and prototyping to production and support.

Have a problem like one of these?

Start a project