GenAI for psychometric assessments that scale beyond the counsellor's calendar

Approach

We started from the existing psychometric instrument, not the model. The counselling team walked us through how they interpret trait clusters, what “good” looks like in a written report, and where a generic LLM answer would feel off-brand or clinically wrong. From that we defined three separate GenAI surfaces — item generation, adaptive scoring commentary, and long-form report drafting — each with its own evaluation rubric, so we could harden one pipeline at a time without touching the validated scoring model underneath.

Solution

We shipped a Next.js assessment portal backed by a FastAPI scoring service. A private item bank feeds a retrieval layer that grounds every generated question in the original construct definitions, and a constrained prompt pipeline drafts counsellor-ready narratives in the practice’s own voice. Counsellors review and edit drafts in an inline editor before a report is finalised, and every approved edit flows back as a preference signal that tunes the next generation. Integrations with the existing proctoring layer keep high-stakes sessions honest without changing the candidate UX.

What changed

Counsellors now walk out of a session with a finished report in hand instead of a backlog to write up that evening. The generative item bank gives the practice room to refresh tests by cohort without commissioning a new instrument every cycle, and the scoring commentary reads like the team wrote it — because, in effect, they did. The measured shift shows up in the outcome metric above.