Andreas Heige

Project 2 of 3 · Live, weekdays

Lunchit

Today's lunch around Platinan in Gothenburg on one page — read from each restaurant's own site, tagged veg, fish or meat, and translated to English.

The idea

I was tired of clicking around several sites every day to see what was for lunch near Platinan, the building by Centralen in Gothenburg where Knowit has its office. The sites that already collect menus did not help. They were dull, and the lunch was buried in noise. Lunchit is the page I wanted instead: today's menus, nothing else, laid out so they are pleasant to read.

It was built for my colleagues. It spread from there, to more of Knowit and to other companies and buildings around Platinan, and it went out in a Knowit newsletter. Colleagues I don't know have given me a pat on the shoulder and said it is fun.

Around the menus

The fun for me is not the product, showing lunch, but everything around it. The site was a playground on purpose, more or less.

Every restaurant writes its menu its own way, so each one gets its own parser. A model on Workers AI's free tier, Llama 3.3 70B, has three narrow jobs: guess veg, fish or meat when the menu has no label, translate the dishes into English, and write up each release for the site's news page. There is a loop in that last one: the site changes, and the model that is part of the site writes down what changed.

Then the two parts I enjoy most. There are stats to experiment with: ratings turned into a league, the menus' ingredients drawn as a graph, a weekend recap of the best-rated and returning dishes. And there is issue reporting. Anyone can report a problem from a form on the page, without a GitHub account. The report becomes an issue, and the person who sent it can ask for an email when it closes.

The struggle

The menus live on other people's sites, and those sites change without notice, each in its own way. Some are plain HTML, some are PDFs exported from Canva, one is built from Wix text blocks. One blocks headless browsers, so a real browser window opens on the build server just to read it. One page lays its week out in two columns, and its source runs column by column, so read line by line, Wednesday came out empty and Thursday showed Wednesday's dishes. The parser now pairs each text block with its day heading.

The work is keeping an app that catches those sudden changes. Every weekday morning the build reads every menu. A fetch that fails gets one retry, and a menu that still cannot be read becomes an issue, unless the live site already has that day's menu. Each parser is tested against a saved copy of its page; when a site changes, the copy is captured again and the parser is fixed to match. The model is checked too: an answer that is not a diet is dropped, a missing translation leaves the Swedish, and a release it fails to write up gets plain fallback text. It still gets things wrong. A vegetarian dish with a meat-sounding name was filed as meat until a fix taught the rules otherwise.

So the model does small jobs inside fences, and the tests, retries and issues do the rest: nothing repairs itself, but when a site changes, someone hears about it before lunch.

None of this code was written by hand. Agents wrote all of it, working from intent I stated and checked: , or IDD. The idea is to make the intent explicit, so the agents do not have to guess it from a short prompt and drift.

  1. restaurant sites HTML, PDF, Canva
  2. one parser per site
  3. Workers AI veg, fish or meat
  4. today's menus cached 30 min
  5. lunchit.se
  6. Workers AI English names
  7. D1 served dishes, stars
  8. stats and weekend recap
  1. restaurant sites HTML, PDF, Canva leads to one parser per site: page text
  2. one parser per site leads to Workers AI veg, fish or meat: unlabelled dishes
  3. lunchit.se leads to Workers AI English names: English mode
  4. lunchit.se leads to D1 served dishes, stars: ratings
  • Workers AI veg, fish or meat leads to today's menus cached 30 min
  • today's menus cached 30 min leads to lunchit.se
  • today's menus cached 30 min leads to D1 served dishes, stars
  • D1 served dishes, stars leads to stats and weekend recap
  • Read
    • restaurant sites HTML, PDF, Canva
    • one parser per site
  • Serve
    • today's menus cached 30 min
    • lunchit.se
    • Workers AI English names
  • Remember
    • D1 served dishes, stars
    • stats and weekend recap
Fig. 1 — From each restaurant's own page to one list, every weekday.
Reads
Each restaurant's site, PDFs too
Tags
Veg, fish, meat, allergens
Model
Workers AI, free tier
Remembers
What was served, and the stars
Every restaurant writes its menu its own way. One parser each, so nobody has to read them all.