Booze deals and cocktails
Alcohol deals across ten supermarket chains, plus cocktails and party games. Web, iOS and Android out of a single monorepo.
Fixuj & Mixuj started in 2021 as an Instagram account reviewing booze and posting cocktail recipes. It moved to TikTok, gathered twenty thousand followers and then went quiet.
The restart came when a parallel account with fifty thousand followers shut down and left the market empty. This time not as content, though, but as a product: an app that actually watches the prices.
2021
Booze reviews and cocktail recipes.
2022
Twenty thousand followers.
2023
The account stopped posting.
2026
This time as a product, not as content.
~990
products in the catalogue
15 subcategories, ~313 brands
10
chains under watch
Kaufland, Lidl, Billa, Tesco, Penny…
3×
price runs a day
05:00 · 12:00 · 19:00 CET
14
days of forecast
p10–p90 quantile band
Cheapest is not the same as best value. Beyond price, the catalogue sorts by discount, by price per litre and by "B4B", meaning how many millilitres of pure alcohol you get per crown. Search tolerates typos and reaches across brands and types.
Web, iOS and Android share one monorepo, so components, queries and types are written once. The native app adds a push when a price drops below a watched threshold. The Games section runs on web and Android; it is hidden in the iOS build, because drinking-game content would not clear App Store review.
Jägermeister 0.7 l, one day, eight chains. Three hundred at Billa, six hundred and five at Košík. This is the whole product in one picture. And exactly why it pays to know when to wait.
Real data from 13 August 2026. Regular price 567 CZK; discounts are calculated against it.
Not because the other road was closed. Because this one worked better and we can say why.
A separate scraper for every retailer.
A single source: kupi.cz.
Six scrapers means six DOM structures that keep changing, six anti-bot defences and six local quirks. The aggregator does that job better than we could maintain it.
Risk: dependency on one source. Knowingly accepted.
Playwright, the way it is usually done.
Plain HTTP and cheerio.
The source returns complete server-rendered HTML, so a browser adds nothing. Without it the run is an order of magnitude faster and the parser can be tested against saved fixtures with no network at all.
A full pass over every product takes 2 to 4 minutes.
One document per product variant, the way indexes are usually built.
One document per variant × store pair.
Filtering to "Billa only" and sorting by price per millilitre of alcohol filtered the right products but ranked them by the best offer across every chain. The card then showed a Penny price and a Penny logo inside a result that was supposed to be from Billa. Collapsing at display time fixes the contradiction in the data instead of patching it in the frontend.
The index grew from about a thousand documents to tens of thousands. Search speed did not notice.
A monorepo: two apps and three shared packages. Forecasting runs on a hosted service; the repo holds the data pipeline and the accuracy scoring.
Scrapers, pipelines, search and native apps are things we build for clients too: web application development