What it is
Diagnostic assessment and student analytics. Next.js, PostgreSQL, containerised test infrastructure with CI. An LLM pipeline turning raw assessment responses into individual reports, with prompt architecture and output-quality evaluation iterated against expert feedback rather than against a metric. 1000+ tagged questions, 3000+ test-takers.
The interesting constraint
The reader of the output is a fifteen-year-old and their teacher. A report that is correct and unusable has failed. Evaluating that is not a benchmark problem, and I do not think we solved it.
To writeWhat you tried: the evaluation approach, what worked, what did not.