Zonnemens.
My roleConcept, brand and platform
Zonnemens, first called N1 Project, is a health community I made from concept to code: the brand, the platform and the AI coach. Members follow lessons per theme, protocols and live challenges, buy tickets for masterclasses and events, and ask the Zonnemens Agent for advice. The coach searches the lessons before it answers, links to its sources and remembers what a member is working on. One codebase for web, iOS and Android.
- What I did
- The concept, the brand and the platform with its AI coach, designed and built with AI.
- Content
- Lars van den Nieuwenhoff: the lessons, protocols and challenges.
- 12
- themes in the university
- 3
- platforms from one codebase: web, iOS, Android
- 5
- protocols, step by step
- RAG
- a coach that answers from the lessons
The home page: what a Zonnemens is, the method, the day clock and the challenges, as a new member sees it.
- University
- Protocols
- Daily practice
- Day clock
- Challenges
- Masterclasses and events
- Community
- Membership
- Zonnemens Agent
- Knowledge
- Shop
- Back office
A health platform with a coach that reads the lessons before it answers, and remembers what you are working on.
- Lessons cut into overlapping chunks
- Each chunk gets a context summary, then an embedding
- Vector, Dutch full-text and trigram search in one SQL function
- Fused with RRF, reranked by a model, six sources at most
- Member context: progress, streaks and goals from a memory file
- Answer posted, every lesson link checked against the database
Contextual retrieval
Before a chunk is embedded, a model writes two sentences on where it sits in the lesson, so short passages still find their source. When the lessons hold too little, the search widens to the community.
A coach that remembers
Every night a model distils lasting facts from the conversations into a dossier per member, goals among them: at most 25, updated or archived as things change. Members can switch it off.
Evals on the real pipeline
A golden set of about fifty questions runs every week through the production pipeline and is judged on grounding, relevance and tone. A member's thumbs-down becomes a new test case.
It learns what works
A weekly agent proposes A/B tests on mail opens and message replies, picks winners with a two-sided z-test and keeps a journal of what it learned. The feed follows a taste vector that also learns from what members skip.
Spaced repetition
Practice is scheduled in the spirit of half-life regression: a right answer doubles the interval from the time that actually passed, a wrong one halves it, and missed questions return in the same session.
Its own sun maths
The day clock runs on its own port of the solar formulas, with an epoch error fixed, and stays within 21 seconds of NOAA over two years and three locations.
Spaced repetition
The real practice player: short reads and questions as multiple choice, fill in the blank or true or false, with an explanation for every answer. A right answer doubles the time to the next review; a wrong one halves it and comes back in the same round.
Five pillars
Everything on the platform hangs off five pillars, such as energy, anatomy and attention. So every lesson shows what it builds on.
On the rhythm of the sun
Its own port of the solar formulas works out, for each day and place, when to catch light, eat, move and wind down. Checked against NOAA: within 21 seconds, over two years and three locations.

Always something to join
Live challenges all year, like a winter challenge or a dopamine reset, with a calendar and a countdown to the next one.

Web and app in one
Paths, lessons, protocols and chats in one React codebase, wrapped as an iOS and Android app with Capacitor. A bar at the bottom for practising on the go.
