I'm Joël. I build simple solutions to complex problems.

Selected work

My years at Gynzy also covered student scoring, CRM integrations, real-time classroom quizzes, APIs for the apps and retiring old systems.

• Gynzy · 2024–2026

  Fixing school login

  On a school visit I saw students struggling to log in. The data showed how big the problem was, so I started and led the move to Basispoort, the national school login, and a direct sync with schools' administration systems (LAS). We went from design to dry run, pilot and full rollout; another team built the auth layer.

  Failed logins dropped sharply, and the LAS connections gave us reliable data on every teacher and student in a school. That helped usage grow and fed the sales funnel.

  • TypeScript
  • NestJS
  • OAuth2 / OIDC
  • Prefect

• Gynzy · 2020–2025

  A data platform that lasted

  I architected Gynzy's data stack after comparing the alternatives. When I left six years later it still ran on the same foundations, with hundreds of tested models.

  • BigQuery
  • dbt
  • Prefect

• Freelance · Stekkies

  From one big server to several countries

  Rental-listings service Stekkies ran on one large VPS. I helped move it to Docker Swarm and then Kubernetes, splitting out separate services. After that it could launch in more countries and languages.

  • Docker Swarm
  • Kubernetes

• Freelance · web data

  500,000 products a day

  I scrape websites for product information and images, turn it into clean, structured data, and deliver it in whatever form a client needs, such as S3 or push. The hard part is routing requests so you don't get blocked, and building auto-healing scrapers that use AI to repair themselves when a site changes. Stichd, an apparel maker for premium brands, has been a client for years. I've also done this for Tommy Hilfiger and for steel processor MCB.

  • Python
  • Prefect
  • dbt

• Gynzy · 2023–2026

  AI at Gynzy

  I helped shape how Gynzy approached AI across the company, from strategy to how we worked day to day.

  In the product

  • An AI quiz generator for teachers.
  • Text-to-speech, replacing a vendor in about two months, with a phonetic rule set to correct wrong pronunciations.
  • Image generation, with a pipeline that keeps images consistent, safe for the classroom and working in any language.
  • Background removal for images, on a model we host ourselves.
  • Semantic search across our library and images.

  In how we work

  • Answering AI questions from across the company.
  • Helping developers build an intuition for why certain prompt techniques do or don't work, drawing on my data science background.
  • Setting out a vision for product development, with discovery as small experiments in the product.
  • Helping colleagues vibe-code where that fits, with guardrails.
  • Python
  • FastAPI
  • Azure OpenAI
  • Imagen
  • Typesense

• Hackathon · 2026

  Hacking for Good

  In one day at a Google hackathon in Ghent, four of us built a working prototype for BSPHO, the network of Belgium's paediatric cancer centres. AI drafts consent forms for clinical studies in three languages and flags what a person needs to check. See the repo.

  • Python
  • TypeScript
  • Vertex AI

• Gynzy · 2019–2024

  A database choice that held

  Student answers peak at 100 a second. We compared SQL, time-series, graph and document stores and chose MongoDB. I ran the first service in shadow for a long time before the team took it to production.

  • NestJS
  • MongoDB