I thought y'all might be interested in a legit "local is better" use case.

TLDR: Qwen 3.8 FN solved a work problem that requires the privacy of a local model, with the same skill as a frontier cloud model.

I run an MSc dissertation module at work which involves tracking students, supervisors, offered projects, project choices, examiners and eventually grades. None of it is super complex, but it's all annoying -- students choose their top five projects from an offered list, but supervisors also choose students; supervisors can't mark their own students work but can mark others etc. The kind of low level matching work that, done manually, fries my brain.

A simple relational database is the obvious solution but, before gen AI, nobody had time to code up a database for processes that only happen once per year. There's also the ingest problem: data comes from ms forms, from the student records database, and from moodle, all with csv export in different formats, with e.g. different arrangement and spelling of middle names etc.

I'm a heavy codex user, but I can't be offering up the names and emails of students and supervisors, nor the details of early-stage research projects, to the cloud.