752 Nigerian institutions, shipped as a repo instead of an endpoint

The hosting decision is the least glamorous thing here and the easiest to defend. The author's stated reasoning is that a single free-tier server cannot reliably serve production traffic for everyone, while a project that is trivial to self-host scales to however many people need it ^[3]. Reference data suits that trade. A list of polytechnics does not change between requests, and the install path is a clone, an npm install and an npm start, with no database server, no config file and no key required by default ^[4]. The whole dataset is three CSV files compiled into a local SQLite database by a prestart script at boot ^[5], which is why the README can offer Render, Railway, Fly.io, Docker or a plain VPS as one-command deploys: there is nothing to provision ^[14].

Now the arithmetic. The project reports 1,700-plus parsed departments ^[9] across 752 institutions ^[1], which averages about 2.3 departments per institution ^[12]. A single federal university carries more than that inside one faculty. So the structure layer is thin in places, and the project says so in the record itself, marking a structure as not clearly itemized in public sources rather than inventing one ^[7].

The course layer is where the honesty is load-bearing. Those rows are not looked up, they are produced: polytechnic departments get the ND to HND ladder, colleges of education get NCE, and university departments get a qualification chosen by matching subject keywords against a documented table, defaulting to B.Sc. Every generated row carries source: "inferred" in the API response, not only in the README ^[8]. Since the verified tier does not exist yet ^[9], every course row in the dataset today is an inference ^[13]. That is a fair deal only if consumers read the field.

The parser bug is the argument for shipping the repo, stated better than any marketing line would. A sanity check on the generator's output turned up strings like "ND Aba" and "HND est. 1992" ^[10], because some source cells tuck an establishment year, a renaming or an ownership note inside the last school's own parentheses instead of using the separator the parser expected ^[11]. The generator inherits whatever text the parser leaves in the department field, so a punctuation convention in a spreadsheet becomes a fake qualification. Against a hosted endpoint, an outsider seeing that would file an issue and wait. With three CSVs in the clone ^[5], they can walk from the bad response back to the offending cell and send a diff.

That is the practical split for anyone building on this. Filter on the inferred tag and you get a smaller dataset you can stand behind. Ignore it and you ship a plausible catalogue of degrees, generated from keywords, that no institution has confirmed ^[8].