2026
Interview Scoring: LLM-assisted HR review
Nhat Nguyen (Nate)
Problem
Hiring teams need to score interview answers consistently without reading every response from scratch. Candidates answer six questions through a personal link. HR should see a suggested score with reasoning, then keep it or override it with a required reason.
What I owned
End-to-end product: a FastAPI API with in-process scoring workers, a Next.js HR app (BFF with httpOnly cookies), and a Next.js candidate app. PostgreSQL is the system of record and the scoring queue. The API runs on a VPS; both web apps run on Vercel.

Architecture
Three independently deployable apps.
The API owns all business rules.
Scoring is an asyncio worker pool inside the API that claims jobs with SELECT ... FOR UPDATE SKIP LOCKED.
A DeepSeek adapter must return {score: 1-4, reasoning}; invalid output is rejected.
Failed jobs never expose a score.
HR sessions use rotating refresh tokens held in httpOnly cookies. Each HR account only sees its own candidates; cross-account access returns 404. Candidates use an unguessable session token and never see scores or grading guides.
Text answers are hashed for idempotency, capped at three changed submissions, then scored.
Audio is stored and shown to HR but is not transcribed in this version (needs_a_person).

Outcome
A live HR demo, not a mock. Roles, candidate links, LLM scores, and required-reason overrides run against the production API.
Try the demo
Open the HR app and sign in with:
- Email:
[email protected] - Password:
4uuXKGhi24zDU4MShW8dpJRK
You can also create your own account from sign-up. The candidate app is a separate surface: HR creates a personal link, the candidate answers, then scores show up for review.
