CrateDigger

In a record shop, when I find a record I don't know, I only have moments to look it up, listen to a few tracks and decide whether it's to my taste. I also need to know if the price is fair, whether I already own it, and where it would sit among the records on my shelf and the music in my library.

Discogs and Apple Music each hold part of that answer. Neither says "you have two records like this already" or "nothing here is close".

I came up with the problem and described the features. I drafted and refined the requirements with an LLM, then handed them to AI coding agents (Claude Code and Codex). The agents wrote all of the code.

My part was directing the work, labelling the evaluation data by hand, testing the suggestions in shops and making the calls.

I scan a barcode, or type a catalogue number or an artist and title. The card then shows whether I own the record, its main style (one of 24 families) and up to three of the closest records from my shelf and my Apple Music library. If nothing is close, it says so. It also plays 30-second previews of three tracks and shows the lowest Discogs price.

Buy adds the release to my Discogs collection and Wishlist adds it to my wantlist. Both can be undone, and every tap is saved in a dig history.

The key decision

My first test, on my own collection, showed that Discogs tags are too coarse to say what sits close to what. So a language model now chooses the matches, but only from records I already have. The app drops any suggestion that is not in my collection or library.

It runs on a private server and is reachable only over a private network. It has no login, so it is never exposed publicly.

I labelled a 20-record test set by hand and ran both models on 18 October 2026. The primary model is the one the app uses. Two records with uncertain reference labels are left out of the style score, which is why its denominator is 18.

Style accuracy came in just under its target. False close matches sat exactly at the limit. I kept the primary model because it made fewer false matches (3 against 5). The decision-in-under-60-seconds target passed in live use in record shops.