Jon Kraayenbrink tagged 479 LinkedIn saves for $0.0195
An open source tool runs Jev over your LinkedIn bookmarks and labels topic, hook and format so you can search them.
Jon Kraayenbrink open sourced a tool that runs Jev over your saved LinkedIn posts and tags each one with a topic, a hook and a format, so the pile becomes searchable. He says it classified 479 posts in 225 seconds for $0.0195. The gap it fills is mundane and real, LinkedIn has no native way to search what you saved, so a bookmark from six months ago is effectively gone.
479 posts classified in 225 seconds for $0.0195
The numbers
225 seconds across 479 posts is about 0.47 seconds per post, and the whole run costs about two cents, which works out to roughly four thousandths of a cent per post. Those are the only figures Kraayenbrink posted. He does not say whether the run was sequential or batched, what hardware or client it ran from, or how large a typical saved post is, so treat the per post latency as a throughput number for one archive rather than a response time you could quote in a product spec.
Three labels per post is also the interesting design choice here. Topic is the obvious one. Hook and format are the kind of soft editorial categories where an embedding index gives you fuzzy neighbours and a classifier gives you a field you can filter on. Kraayenbrink's framing is save, classify, search, which is a pipeline that only works if the labels are consistent enough to filter by.
What is not measured
There is no accuracy figure. Nobody has posted a count of how many of the 479 posts got the right topic, hook or format, and there is no held out set, no second rater, and no comparison against an embedding based search over the same archive. A personal bookmark search is forgiving of mislabels in a way that a production classifier is not, because the person running it wrote the taxonomy and can eyeball the output. That is a legitimate place to ship without an eval. It is not evidence about label quality on anyone else's corpus.
For people building on it, the economics here are a batch job, not a hot path. Half a second per item is fine when you run it once over an archive and then query the results, and it is not fine when it sits between a user and a page load. The two cent total is for 479 items in one pass, so the figure to carry forward is cost per item at that size, not the headline.