Dev log

How AllCars scores 11,000 Cyprus used cars in under 90 seconds

You're looking at a 2018 hatchback for €12,500. Is that a deal, a normal price, or a "drive an hour to be disappointed" situation? Here's the human version of how AllCars answers that question across every used car in Cyprus.

Buying a used car in Cyprus shouldn't feel like detective work, but it usually does. You scroll, you compare a few similar listings, you guess. Maybe you spreadsheet. Mostly you just hope. AllCars exists to take the guessing out — for every car in the index, I work out where it sits against the rest of the market and stick a deal score on it.

Sounds simple. The first version of my scoring engine couldn't do it fast enough, though. It scored cars one at a time, hitting the database for each one, and a full rescore took most of an evening. By the time it finished, the market had already moved a bit. So I rebuilt it.

From 11,000 little queries to three big ones

The new engine doesn't ask the database eleven thousand questions. It asks three: give me the active listings, give me the segment summaries, give me the comparable set. It loads all of that into memory, and from there it's pure number-crunching. The database does the boring "fetch a lot of rows" thing it's already good at, and the scorer never has to wait for a round trip.

That single shape change is most of the story. Per-listing logic became a vector operation. Sequential became parallel-friendly. And the engine started behaving like a calculator, not a chatty client.

Why "similar cars" is harder than it sounds

Comparing two cars feels obvious until you try to write it down. Make and model are categorical. Year and mileage are numeric. Body type matters but not as much as engine size. Colour barely matters. How do you get all that into one number?

The trick is something called Gower distance. It treats every feature on its own scale and blends the results, so categorical and numeric features can co-exist without one drowning the other out. Combined with a smooth weighting curve (Epanechnikov, if you want to look it up), each comparable contributes proportionally to how similar it really is — not "in or out" by some arbitrary threshold.

The result: a deal score from 0 to 100 that holds up across the full range, from common Toyotas with hundreds of comparables to one-of-a-handful imports with five.

Why this matters when you're buying

For you, the buyer, all of this is invisible — you just see a number next to a listing. "Steal" at 90+. "Good Value" at 75–89. "Average" in the middle. What changed under the hood is that the number is now actually fresh: I rescore the whole market in roughly a minute and a half, multiple times a day. The car you saw priced fairly this morning is still scored against the market right now, not the market from last week.

If you've ever spent a Saturday driving across the island to view a "great price" that turned out to be a normal price for a tired car, this is the thing that's supposed to save you that Saturday.