Blog/Private AI for technical interview question banks
Hiring

Private AI for technical interview question banks

By Samuel Seidel · Updated September 9, 2026

A technical interview question bank works only while candidates don't know it. Generating it with a general-purpose AI tool works fine the first time; the problem is the copy of the bank left in that vendor's logs afterward. A generation model on a dedicated machine leaves no such copy in a shared service.

Why interview questions leak faster than most internal documents

Question banks circulate more widely than most people expect: candidates share them on interview-prep forums, engineers post them to private group chats after a loop, and some candidates paste questions into their own AI tools mid-interview to get an answer. None of that is new to AI-assisted hiring. What's specific to AI-drafted question banks is the extra copy created at generation time, sitting on infrastructure the hiring team doesn't control, under a retention policy nobody on the hiring team reviewed.

There's also a subtler risk with shared, general-purpose models: a phrasing or a specific coding problem generated for one company's loop can resemble what the same model generates for a different company asking a similar prompt, because both requests draw from overlapping patterns the model has seen. That's not a data breach in the traditional sense, but it does mean a hiring team asking a shared service for "a medium-difficulty graph problem for a backend interview" isn't necessarily getting something unique to their own process. Self-hosting doesn't fix that part, since open models share training patterns too; feeding the model your own past questions and asking for variations on them does more.

What changes when the generating model is self-hosted

Running the question-generation assistant on a dedicated DGX Spark keeps the prompts, the drafts and the final bank on one machine assigned to the hiring team, with no chat-product log of what was asked or generated. The recruiting team still gets a model that can draft coding problems at a target difficulty, or vary an existing question so it can't be matched against a public writeup.

A practical setup is Open WebUI, restricted to the hiring managers and recruiters who need it, with prior loops and rubrics loaded as reference material (the approach in the on-premise RAG guide). New questions can then be checked against what's already been used. If you review candidates' take-home submissions with AI, the same logic applies; the coding agents doc shows how to run one against a private model.

What the model is useful for, and what stays with the interviewer

A generation model is a reasonable help for drafting question variations, checking a rubric against a candidate's written answer, and keeping a bank of questions organized by difficulty and topic. The hiring decision itself, and any judgment call on a borderline or unconventional answer, should stay with the human interviewer. That's partly a fairness question and partly a legal one: several jurisdictions already restrict how much weight an automated system can carry in an employment decision.

What this doesn't solve

Self-hosting the generation model doesn't stop candidates from sharing questions with each other after the fact, and it doesn't replace a hiring team's own process for rotating stale questions out of the bank. What it removes is one specific exposure: an AI vendor holding a copy of your question bank before you've decided who gets asked what. See pricing for what a dedicated Spark costs.

Related pages

First top-up: pay $20, get $30 in credit

Keep your question bank out of shared AI tools.

A dedicated Spark for hiring, starting at $0.79/hour, deployed in minutes from EU-Central. Read the Open WebUI setup →

Your data stays on a machine assigned only to you. GPUwerk processes it under its published DPA.

Create account See private LLM hosting