An AI recommendation engine that turns career counselling into a product

Approach

We started by treating the counsellor, not the student, as the primary user of the model. Counsellors already had a mental schema for how an aptitude profile, a set of interest signals, and hands-on experiential scores map to career families — so we encoded that schema explicitly instead of letting a general-purpose LLM guess at it. Student signals became structured features; career paths became retrievable documents with attributes like required aptitudes, day-in-the-life tasks, and entry pathways. The LLM’s job was ranking and explanation, not invention.

Solution

We shipped a two-stage recommendation service behind a FastAPI backend. Stage one is a deterministic scorer that matches a student’s assessment vector against a pgvector index of career profiles and returns a ranked shortlist with feature-level contributions. Stage two passes that shortlist — plus the student’s experiential-session notes — to an LLM that writes a counsellor-voiced rationale, flags mismatches the scorer surfaced, and suggests next experiential modules to try. Counsellors see the draft inside an editing UI, tweak or reject sections, and publish. Every edit is logged as a training signal for prompt and retrieval tuning.

What changed

The counsellor workflow flipped from “write the report” to “review and shape the draft,” which is where their judgement actually adds value. Turnaround on a personalised recommendation report went from a multi-day cycle to minutes of counsellor editing time, and the platform now has a live product surface — not a PDF — that students return to as their interests evolve.