How would you build software like this without RDF and the Semantic Web?

The trouble is that the evolution of OWL, RDFS and SPARQL stopped before they reached the useful threshold.

For instance there is the subproprerty relationship that would let you map temperatures from european sources to temperatures from US source if and only if you could specify a conversion formula from centigrade to Fahrenheit.

Some kind of of universal namespace is necessary for linking data is necessary, but it cannot be

   http://some.url/accessible/to/everyone/and/frozen/in/time

for many reasons, one of which is that there is no standard way to relate this to

   https://some.url/accessible/to/everyone/and/frozen/in/time

and also that not everyone who wants to refer to a concept wants to a mint a URL. For one thing the data might be private and no dereferencing is allowed. For another thing I might be able to afford to upload a Turtle file to a web site but that doesn't mean I can afford to run a deferencing server or public SPARQL endpoints.

Many scaling properties of real data sets make both the dereferencing server and the SPARQL server untenable. Often there is one ?p ?o such at

   ?s ?p ?o .

holds for 1/3 of ?s. Because of these unbalances you will fail to dereference some concepts, have some SPARQL comments time out, generally get wrong answers where you might not know you got a wrong answer and not have a clear approached (reified by standards) to get a right answer.

So you see a lot of "Linked Data" demos that insult the intelligence of the viewer because they smush together a lot of wrong stuff and the result looks like that trash compactor for Star Wars but the people pushing them just don't listen to the criticism that... look, dude, 30% of the facts here are wrong!