# My friend asked for interview practice, so I built an AI interviewer that runs locally

> Source: <https://dev.to/abhishek_gupta_982095928c/my-friend-asked-for-interview-practice-so-i-built-an-ai-interviewer-that-runs-locally-2obb>
> Published: 2026-10-04 18:07:57+00:00

*This is a submission for the [Hacktoberfest Weekend Challenge: Build for a Friend](https://dev.to/challenges/hacktoberfest-weekend-2026-10-01)*

FRIEND is in placement season and told me one thing scares them most: SPECIFIC FEAR, e.g. "I freeze when they ask about my own project." Paid mock interview tools cost money and want a full resume uploaded to someone else's server.

So I built **Mock Interview Partner**: upload a resume, pick a role, and it interviews you. It asks questions about your actual projects, pushes back on vague answers with follow-ups, scores each answer strictly from 1 to 5, and writes a report with weaknesses and a practice plan. It runs completely on a laptop.

(Optional: a 30 to 60 second screen recording.)

A free, private mock interviewer for placement season. Upload a resume, choose a target role, and get tailored questions, follow-ups, strict per-answer feedback and a final report. It runs entirely on your own laptop with an open-weight model, so nothing is sent to a server.

Built for the Hacktoberfest Weekend Challenge: **Build for a Friend**.

A resume contains a phone number, address, grades and email, and a mock interview records someone's honest mistakes. Neither belongs on a server you don't control. Running a local open-weight model means nothing leaves the laptop, and no internet is needed once the model is downloaded.

Open source also gives you:

`llama3.2:3b` through Ollama, fully on a 6 GB RTX 4050
The flow is: resume text → structured profile → question plan → answer → score and feedback → optional follow-up → final report.

Everything behavioural lives in plain text: the prompts, three interviewer personas and a `rubric.yaml` with score anchors.

**What went wrong:** WRITE ONE REAL FAILURE. Ideas: the model gave 4 or 5 out of 5 to weak answers until I added explicit score anchors and told it to grade like a strict interviewer. Or the 7B model split between GPU and CPU on 6 GB of VRAM and was painfully slow, so I moved to the 3B model (X seconds per answer instead of Y).

I ran a mock interview with FRIEND_NAME, who is preparing for placements, and asked for honest feedback afterwards.

Their first reaction was a simple **"wow."**

What they found most useful was the **improved sample answers and the review** after each question. When I asked what the tool showed them that they hadn't noticed, they said: *"I missed too many things, and this app pointed them out. I improved my answer."*

I also asked whether any scoring felt unfair, since a strict grader is only useful if it's fair. They said no, the feedback looked fair to them.

Their verdict on using it again before placements: *"Yes, it will help me a lot."*

That last answer is what I built it for: a free place to practise as many times as they need, without a resume or answers going to anyone else's server.