Projects

What I build on my own time. The common thread is guardrails: making a system prove its own claims before I trust it.

The five below are private, either because they hold my own health and money data or because they are not finished. Happy to walk through any of them. The ones further down are public and linked.

pulse

Next.js • TypeScript • Postgres (Neon) • intervals.icu • COROS MCP

A training, sleep and recovery app for one athlete, me, installed on my phone. It brings together what lived in four places: Garmin Connect, COROS, Strava and intervals.icu.

Readiness is a traffic light, not a 0 to 100 score. It compares the last seven nights of HRV with my own 60-day normal, after published methods, and resting heart rate, sleep and form can only flag, never decide. Every model names its method and its limits in the source file. Each forecast was back-tested against "no change" before it was shown, and it comes as a range sized from its past errors.


Repères 2027

Astro • TypeScript • no client dependencies

A voting advice app for the 2027 French presidential election. In this category credibility is the product, so the build enforces it. Every candidate position needs a dated source, every candidate answers the identical set of statements, and the build fails if either is untrue. Ties are broken by a session seed and shown as ties. There is no server, so answers never leave the browser.

Each rule answers a failure of Elyze, the app that did this in 2022. It broke ties by internal candidate id, so the same candidate always won. It covered a fraction of the declared candidates. And it reserved the right to resell user answers, until the CNIL announced it would check and it stopped collecting data. Ships around 10 KB of JavaScript.


Wattson and Sparfuchs

Python • Claude Code skills

Two personal advisors built the same way. Wattson is an endurance coach that reads my training data from intervals.icu. Sparfuchs handles tax and money as a French citizen living in Germany.

The architecture is the point. Scripts do the calculations and hand back JSON; the model reads it and gives an opinion. In Sparfuchs every legal constant sits in one file with its source and the date it was checked, and the setup script complains when that date gets old. It is my working answer to where a language model is allowed to be trusted, and the honest limit is that the split is a convention I hold myself to, not something the code enforces.


llm-vps

Terraform • Docker Compose • llama.cpp • Caddy

A self hosted, OpenAI compatible LLM endpoint on a rented box, defined as code. Point any OpenAI SDK at it and it works.

The README opens with a cost table, and further down it concedes the thing a README like this usually leaves out: for small, bursty workloads a hosted API running the same open weight models is cheaper and much faster. Self hosting wins on fixed cost, data locality and no rate limits. Pick it for those reasons, not price.


Portulan

Next.js • TypeScript • Postgres (Neon) • Vercel

A quote and proposal tool for small French travel agencies. One structured trip produces the margin sheet, the branded PDF and the client's trip page. It is live, with a first agency using it and a self-serve free trial.

The success bar was set by the design partner, not by me: a real quote sent without Word, then a second trip built from a hotel she described in the first. The AI features follow the same rule. A supplier email read into cost lines, or a hotel filled from its own web page, is a draft she rereads before anything is saved, and the client's contact details are taken out of a supplier email before it goes to a model.

merk

Next.js • Postgres (Neon) • Drizzle • Better Auth • FSRS

A language course and flashcard app I use to learn German (A1 to B2) and Moroccan Darija, installable on a phone. Lessons, part exams, Goethe-style level exams and a placement test, with around 4,600 German words scheduled by FSRS.

The AI teacher gives feedback on writing and spoken answers, but it is fenced in: daily limits per person, a cache, a monthly budget on the gateway, and a fallback to marking yourself when it is off. Every lesson sentence is recorded once and served as static files, so audio costs nothing per play.

merk on GitHub


vo2

Next.js • Postgres (Neon) • FSRS • Crossref

Endurance sports science (running, cycling, triathlon) at master's depth, learned one multiple-choice question at a time. Early: the format and the first module are being built.

There is no AI in the app, and every key claim has to carry a source and an evidence grade. A check fails if a cited DOI does not exist in Crossref or a claim has no source. You also say how sure you are when you answer, and confident wrong answers come back first.

vo2 on GitHub


EuroLens

Next.js • TypeScript • Supabase • public EU data only

Track what the European Parliament is voting on, in plain English, and see how your own MEPs voted. Roll-call results for all 720 members, filterable by country and political group.

It started with a language model writing the explanations. I took it out. Explanations are now composed from the official record through a fixed glossary, which means they are instant, reproducible, and cannot invent a fact that is not in the data. On a tool whose whole claim is being non partisan, a model that occasionally makes something up is not a feature. No API keys, no per request cost.

EuroLens on GitHub · Live site


HBO Max x SensCritique

Chrome extension • JavaScript

Puts SensCritique ratings on HBO Max tiles, because I kept opening a second tab to decide what to watch.

The extension on GitHub


Voyages

Next.js • TypeScript • Tailwind • shadcn/ui

A travel agency site with a quote request flow and seasonal SEO content.

Voyages on GitHub