PETER BRAENDGAARD MADSEN ★

… Click here for my CV →

Everything below is running right now. The numbers on these cards are read live from the apps themselves — this page is not a set of screenshots, it is a dashboard pointed at its own systems.

DEMO open it and use it — seeded data, reset nightly

LIVE APP a real app with real users, not a sandbox

NOT PUBLIC runs on my machine, on purpose

IDEA DASH

DEMO

…

reading the board

A drawer that sits in front of the kanban board.

OPEN →

KANBAN

DEMO

…

reading the board

A personal task board. Columns, cards, drag and drop.

OPEN →

INVEST

DEMO

…

reading the dashboard

A portfolio dashboard with a price robot — real prices, invented holdings.

OPEN →

MACRO TRACKO

LIVE APP

EAT

“I ate two eggs and a banana”

A nutrition tracker you talk to in ordinary language.

OPEN →

PODSCRIPT

NOT PUBLIC

TXT

podcast episode → clean markdown

Podcast episodes turned into transcripts worth keeping.

PODDASH

NOT PUBLIC

PLAY

eleven shows, one feed

One place to go when you want to listen to a podcast.

This is the box I actually use. Type something and it is filed into the demo idea board above; → TASK sends it to the kanban board instead. Both are wiped nightly.

FILMEU PARTNER HOLIDAY CALENDAR LIVE APP

When eight European film schools are closed, in one calendar. Built for the FilmEU alliance and handed over.

OPEN →
LIVE APP

What problem does this solve?

Eight film schools in eight countries run FilmEU as one alliance, and none of them could tell when the others were shut. Good Friday is a public holiday in Denmark and an ordinary working day in Ireland — where the film school closes anyway. Short of emailing someone and waiting, there was no way to know that before booking the meeting.

How does it solve it?

Two kinds of day, handled two ways. I wrote a script that calculates the public holidays from each country's own rules — 311 days across three years, including Orthodox Easter for Bulgaria and the weekend-substitution rules Bulgaria and Ireland use. Nobody types a date in, so nothing can be a typo.

School closures cannot be calculated; no rule produces them. Those are researched one at a time, and every entry carries a source you can click. Then it leaves the website entirely and publishes a calendar feed — the ordinary kind Outlook and Google already understand.

Why does it work?

It asks nothing of anyone. No new tool, no login, no system for an IT department to approve or maintain. You subscribe once and the days appear in the Outlook these schools already live in. It adds nothing to anyone's stack — it uses the calendar they already have open all day.

It is allowed to be incomplete. A wrong closure is worse than a missing one, because somebody books a flight around it. So a day nobody can confirm is left out, and an entry that came from a corridor conversation says so instead of implying it was checked. Being trusted is the product.

It was built to be left behind. I handed it over when I left the university that runs it. The person maintaining it now adds a closure through a web form and never opens the code.

Cloudflare Workers · D1 · iCalendar feed · no framework
NOT PUBLIC
The Poddash dashboard: a header counting new episodes, shows and queue, a row of filters, and a list of episodes each with resume, queue and restart buttons, above a player bar.

Tap the screenshot to see it full size

What problem does this solve?

Having multiple apps had me forgetting which podcasts I followed. Shows went unheard by accident.

How does it solve it?

So I built a website I can play all my podcasts from. It has a PWA layer, so on a phone it installs to the home screen and behaves like a regular iPhone app.

Why can't I click it?

Heavily protected, due to questions about rights. Most of the shows are ordinary public RSS feeds — no question at all. A few are not: reaching those uses my own paid subscription, and sharing access would breach the terms I agreed to. So nobody gets a login.

Cloudflare Workers · D1 · PWA · no framework
DEMO

What problem does this solve?

A backlog you throw everything into becomes a graveyard. Ideas are maybes and tasks are commitments, and when they share one list the maybes bury the things you actually agreed to do.

How does it solve it?

Ideas get their own board in front of the kanban, with a one-way door into it. An idea starts as a spark, becomes developing if it survives a second look, and is then promoted into a real task, parked, or killed. Killed ideas stay in the archive — what you decided against is worth keeping.

Why build it at all?

Because the alternative was discipline, and discipline runs out. Everything went into one list, the list got long, and a long list stops being read. A second board is a structural fix for something willpower kept failing at.

Cloudflare Workers · D1 · vanilla JavaScript, no build step
DEMO

What problem does this solve?

Provides structure and a way to prioritize many different projects and everyday things to get done. Also provides a place for research, models, and notes in the “workspace” function.

How does it solve it?

Columns, cards, drag and drop, and then the part that matters: a card opens into a full page. Markdown with checkboxes you can tick, pasted images, Excalidraw drawings and mermaid diagrams.

Why build it at all?

It started as a practice project and stopped being one when other things came to depend on it — Idea Dash promotes into it, and the quick capture box above writes to it. The Kanban board provides structure.

React · TypeScript · Tailwind · @dnd-kit
LIVE APP

What problem does this solve?

Every macro tracker requires you to weigh everything you eat. That complicates the process when weighing is not possible, or when a meal is made of many different ingredients. The tedious work is what kept me from tracking my diet at all. So I made a compromise: trade precision for convenience.

How does it solve it?

You write what you ate the way you would say it — “I ate two eggs and a banana” — and the AI's only job is working out which foods those are and roughly how much.

It never invents a number. The nutrition figures come from Frida, the Danish national food database from DTU Fødevareinstituttet, and that lookup is deterministic: the same food returns the same values every time. The model reads the sentence; the database does the arithmetic. That split is the whole design — a language model guessing calorie counts would be worse than useless, because it would be confidently wrong.

Why give up the precision?

Because the precision was never where the accuracy came from. A weighed banana is 118g and a guessed one is 120 — across a week that gap is noise. The error that actually costs you is the meal you never logged because weighing it was too much bother. A tracker you use every day and that is roughly right beats a precise one you abandon on Thursday.

React · TypeScript · Supabase · Gemini Flash · Cloudflare Pages
Nutrition data: Frida, DTU Fødevareinstituttet

NOT PUBLIC

What problem does this solve?

Podcasts hold a lot of what I want to remember, and none of it is searchable. You cannot quote an episode, cannot paste it into anything, and cannot find the bit you half-remember from six weeks ago. Listening again at 1.5× is not research.

How does it solve it?

Paste a link — Spotify, Apple, an RSS feed — or drop an audio file. It pulls the audio down, transcribes it with Groq Whisper, and has Claude put the structure back: paragraphs, speakers, headings.

What comes out is a markdown file, and that is the point. Markdown drops straight into Obsidian, where an episode links to everything else in the vault, or goes to NotebookLM as a source you can ask questions of. The episode stops being an hour of audio and becomes text you can search, quote and connect to something else.

Why can't I click it?

Because every run downloads audio and spends API credit. A public version would mean paying for strangers' transcripts and hosting other people's copyrighted audio, so it runs on my desktop and nowhere else. That is a decision, not an unfinished feature — and the alternative, a demo that quietly did nothing, would have been worse than an honest closed door.

Python · Flask · yt-dlp · ffmpeg · Groq Whisper · Claude · SQLite